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 intializations 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/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
191 * Bufctl's are used for linking objs within a slab
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 typedef unsigned int kmem_bufctl_t
;
208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list
;
222 unsigned long colouroff
;
223 void *s_mem
; /* including colour offset */
224 unsigned int inuse
; /* num of objs active in slab */
226 unsigned short nodeid
;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head
;
247 struct kmem_cache
*cachep
;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount
;
267 unsigned int touched
;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init
{
283 struct array_cache cache
;
284 void *entries
[BOOT_CPUCACHE_ENTRIES
];
288 * The slab lists for all objects.
291 struct list_head slabs_partial
; /* partial list first, better asm code */
292 struct list_head slabs_full
;
293 struct list_head slabs_free
;
294 unsigned long free_objects
;
295 unsigned int free_limit
;
296 unsigned int colour_next
; /* Per-node cache coloring */
297 spinlock_t list_lock
;
298 struct array_cache
*shared
; /* shared per node */
299 struct array_cache
**alien
; /* on other nodes */
300 unsigned long next_reap
; /* updated without locking */
301 int free_touched
; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache
*cache
,
314 struct kmem_list3
*l3
, int tofree
);
315 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
317 static int enable_cpucache(struct kmem_cache
*cachep
);
318 static void cache_reap(struct work_struct
*unused
);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline
int index_of(const size_t size
)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size
)) {
336 #include "linux/kmalloc_sizes.h"
344 static int slab_early_init
= 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3
*parent
)
351 INIT_LIST_HEAD(&parent
->slabs_full
);
352 INIT_LIST_HEAD(&parent
->slabs_partial
);
353 INIT_LIST_HEAD(&parent
->slabs_free
);
354 parent
->shared
= NULL
;
355 parent
->alien
= NULL
;
356 parent
->colour_next
= 0;
357 spin_lock_init(&parent
->list_lock
);
358 parent
->free_objects
= 0;
359 parent
->free_touched
= 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache
*array
[NR_CPUS
];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount
;
389 unsigned int buffer_size
;
390 u32 reciprocal_buffer_size
;
391 /* 3) touched by every alloc & free from the backend */
393 unsigned int flags
; /* constant flags */
394 unsigned int num
; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder
;
400 /* force GFP flags, e.g. GFP_DMA */
403 size_t colour
; /* cache colouring range */
404 unsigned int colour_off
; /* colour offset */
405 struct kmem_cache
*slabp_cache
;
406 unsigned int slab_size
;
407 unsigned int dflags
; /* dynamic flags */
409 /* constructor func */
410 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
412 /* 5) cache creation/removal */
414 struct list_head next
;
418 unsigned long num_active
;
419 unsigned long num_allocations
;
420 unsigned long high_mark
;
422 unsigned long reaped
;
423 unsigned long errors
;
424 unsigned long max_freeable
;
425 unsigned long node_allocs
;
426 unsigned long node_frees
;
427 unsigned long node_overflow
;
435 * If debugging is enabled, then the allocator can add additional
436 * fields and/or padding to every object. buffer_size contains the total
437 * object size including these internal fields, the following two
438 * variables contain the offset to the user object and its size.
444 * We put nodelists[] at the end of kmem_cache, because we want to size
445 * this array to nr_node_ids slots instead of MAX_NUMNODES
446 * (see kmem_cache_init())
447 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
448 * is statically defined, so we reserve the max number of nodes.
450 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
452 * Do not add fields after nodelists[]
456 #define CFLGS_OFF_SLAB (0x80000000UL)
457 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459 #define BATCHREFILL_LIMIT 16
461 * Optimization question: fewer reaps means less probability for unnessary
462 * cpucache drain/refill cycles.
464 * OTOH the cpuarrays can contain lots of objects,
465 * which could lock up otherwise freeable slabs.
467 #define REAPTIMEOUT_CPUC (2*HZ)
468 #define REAPTIMEOUT_LIST3 (4*HZ)
471 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
472 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
473 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
474 #define STATS_INC_GROWN(x) ((x)->grown++)
475 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
476 #define STATS_SET_HIGH(x) \
478 if ((x)->num_active > (x)->high_mark) \
479 (x)->high_mark = (x)->num_active; \
481 #define STATS_INC_ERR(x) ((x)->errors++)
482 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
483 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
484 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
485 #define STATS_SET_FREEABLE(x, i) \
487 if ((x)->max_freeable < i) \
488 (x)->max_freeable = i; \
490 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
491 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
492 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
493 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
495 #define STATS_INC_ACTIVE(x) do { } while (0)
496 #define STATS_DEC_ACTIVE(x) do { } while (0)
497 #define STATS_INC_ALLOCED(x) do { } while (0)
498 #define STATS_INC_GROWN(x) do { } while (0)
499 #define STATS_ADD_REAPED(x,y) do { } while (0)
500 #define STATS_SET_HIGH(x) do { } while (0)
501 #define STATS_INC_ERR(x) do { } while (0)
502 #define STATS_INC_NODEALLOCS(x) do { } while (0)
503 #define STATS_INC_NODEFREES(x) do { } while (0)
504 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
505 #define STATS_SET_FREEABLE(x, i) do { } while (0)
506 #define STATS_INC_ALLOCHIT(x) do { } while (0)
507 #define STATS_INC_ALLOCMISS(x) do { } while (0)
508 #define STATS_INC_FREEHIT(x) do { } while (0)
509 #define STATS_INC_FREEMISS(x) do { } while (0)
515 * memory layout of objects:
517 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
518 * the end of an object is aligned with the end of the real
519 * allocation. Catches writes behind the end of the allocation.
520 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
522 * cachep->obj_offset: The real object.
523 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
524 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
525 * [BYTES_PER_WORD long]
527 static int obj_offset(struct kmem_cache
*cachep
)
529 return cachep
->obj_offset
;
532 static int obj_size(struct kmem_cache
*cachep
)
534 return cachep
->obj_size
;
537 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
539 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
540 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
541 sizeof(unsigned long long));
544 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
546 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
547 if (cachep
->flags
& SLAB_STORE_USER
)
548 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
549 sizeof(unsigned long long) -
551 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
552 sizeof(unsigned long long));
555 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
557 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
558 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
563 #define obj_offset(x) 0
564 #define obj_size(cachep) (cachep->buffer_size)
565 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
566 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
572 * Do not go above this order unless 0 objects fit into the slab.
574 #define BREAK_GFP_ORDER_HI 1
575 #define BREAK_GFP_ORDER_LO 0
576 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
579 * Functions for storing/retrieving the cachep and or slab from the page
580 * allocator. These are used to find the slab an obj belongs to. With kfree(),
581 * these are used to find the cache which an obj belongs to.
583 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
585 page
->lru
.next
= (struct list_head
*)cache
;
588 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
590 page
= compound_head(page
);
591 BUG_ON(!PageSlab(page
));
592 return (struct kmem_cache
*)page
->lru
.next
;
595 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
597 page
->lru
.prev
= (struct list_head
*)slab
;
600 static inline struct slab
*page_get_slab(struct page
*page
)
602 BUG_ON(!PageSlab(page
));
603 return (struct slab
*)page
->lru
.prev
;
606 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
608 struct page
*page
= virt_to_head_page(obj
);
609 return page_get_cache(page
);
612 static inline struct slab
*virt_to_slab(const void *obj
)
614 struct page
*page
= virt_to_head_page(obj
);
615 return page_get_slab(page
);
618 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
621 return slab
->s_mem
+ cache
->buffer_size
* idx
;
625 * We want to avoid an expensive divide : (offset / cache->buffer_size)
626 * Using the fact that buffer_size is a constant for a particular cache,
627 * we can replace (offset / cache->buffer_size) by
628 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
630 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
631 const struct slab
*slab
, void *obj
)
633 u32 offset
= (obj
- slab
->s_mem
);
634 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
638 * These are the default caches for kmalloc. Custom caches can have other sizes.
640 struct cache_sizes malloc_sizes
[] = {
641 #define CACHE(x) { .cs_size = (x) },
642 #include <linux/kmalloc_sizes.h>
646 EXPORT_SYMBOL(malloc_sizes
);
648 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
654 static struct cache_names __initdata cache_names
[] = {
655 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
656 #include <linux/kmalloc_sizes.h>
661 static struct arraycache_init initarray_cache __initdata
=
662 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
663 static struct arraycache_init initarray_generic
=
664 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
666 /* internal cache of cache description objs */
667 static struct kmem_cache cache_cache
= {
669 .limit
= BOOT_CPUCACHE_ENTRIES
,
671 .buffer_size
= sizeof(struct kmem_cache
),
672 .name
= "kmem_cache",
675 #define BAD_ALIEN_MAGIC 0x01020304ul
677 #ifdef CONFIG_LOCKDEP
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
686 * We set lock class for alien array caches which are up during init.
687 * The lock annotation will be lost if all cpus of a node goes down and
688 * then comes back up during hotplug
690 static struct lock_class_key on_slab_l3_key
;
691 static struct lock_class_key on_slab_alc_key
;
693 static inline void init_lock_keys(void)
697 struct cache_sizes
*s
= malloc_sizes
;
699 while (s
->cs_size
!= ULONG_MAX
) {
701 struct array_cache
**alc
;
703 struct kmem_list3
*l3
= s
->cs_cachep
->nodelists
[q
];
704 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
706 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
709 * FIXME: This check for BAD_ALIEN_MAGIC
710 * should go away when common slab code is taught to
711 * work even without alien caches.
712 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
713 * for alloc_alien_cache,
715 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
719 lockdep_set_class(&alc
[r
]->lock
,
727 static inline void init_lock_keys(void)
733 * 1. Guard access to the cache-chain.
734 * 2. Protect sanity of cpu_online_map against cpu hotplug events
736 static DEFINE_MUTEX(cache_chain_mutex
);
737 static struct list_head cache_chain
;
740 * chicken and egg problem: delay the per-cpu array allocation
741 * until the general caches are up.
751 * used by boot code to determine if it can use slab based allocator
753 int slab_is_available(void)
755 return g_cpucache_up
== FULL
;
758 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
760 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
762 return cachep
->array
[smp_processor_id()];
765 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
768 struct cache_sizes
*csizep
= malloc_sizes
;
771 /* This happens if someone tries to call
772 * kmem_cache_create(), or __kmalloc(), before
773 * the generic caches are initialized.
775 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
777 while (size
> csizep
->cs_size
)
781 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
782 * has cs_{dma,}cachep==NULL. Thus no special case
783 * for large kmalloc calls required.
785 #ifdef CONFIG_ZONE_DMA
786 if (unlikely(gfpflags
& GFP_DMA
))
787 return csizep
->cs_dmacachep
;
789 return csizep
->cs_cachep
;
792 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
794 return __find_general_cachep(size
, gfpflags
);
797 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
799 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
803 * Calculate the number of objects and left-over bytes for a given buffer size.
805 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
806 size_t align
, int flags
, size_t *left_over
,
811 size_t slab_size
= PAGE_SIZE
<< gfporder
;
814 * The slab management structure can be either off the slab or
815 * on it. For the latter case, the memory allocated for a
819 * - One kmem_bufctl_t for each object
820 * - Padding to respect alignment of @align
821 * - @buffer_size bytes for each object
823 * If the slab management structure is off the slab, then the
824 * alignment will already be calculated into the size. Because
825 * the slabs are all pages aligned, the objects will be at the
826 * correct alignment when allocated.
828 if (flags
& CFLGS_OFF_SLAB
) {
830 nr_objs
= slab_size
/ buffer_size
;
832 if (nr_objs
> SLAB_LIMIT
)
833 nr_objs
= SLAB_LIMIT
;
836 * Ignore padding for the initial guess. The padding
837 * is at most @align-1 bytes, and @buffer_size is at
838 * least @align. In the worst case, this result will
839 * be one greater than the number of objects that fit
840 * into the memory allocation when taking the padding
843 nr_objs
= (slab_size
- sizeof(struct slab
)) /
844 (buffer_size
+ sizeof(kmem_bufctl_t
));
847 * This calculated number will be either the right
848 * amount, or one greater than what we want.
850 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
854 if (nr_objs
> SLAB_LIMIT
)
855 nr_objs
= SLAB_LIMIT
;
857 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
860 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
863 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
865 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
868 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
869 function
, cachep
->name
, msg
);
874 * By default on NUMA we use alien caches to stage the freeing of
875 * objects allocated from other nodes. This causes massive memory
876 * inefficiencies when using fake NUMA setup to split memory into a
877 * large number of small nodes, so it can be disabled on the command
881 static int use_alien_caches __read_mostly
= 1;
882 static int __init
noaliencache_setup(char *s
)
884 use_alien_caches
= 0;
887 __setup("noaliencache", noaliencache_setup
);
891 * Special reaping functions for NUMA systems called from cache_reap().
892 * These take care of doing round robin flushing of alien caches (containing
893 * objects freed on different nodes from which they were allocated) and the
894 * flushing of remote pcps by calling drain_node_pages.
896 static DEFINE_PER_CPU(unsigned long, reap_node
);
898 static void init_reap_node(int cpu
)
902 node
= next_node(cpu_to_node(cpu
), node_online_map
);
903 if (node
== MAX_NUMNODES
)
904 node
= first_node(node_online_map
);
906 per_cpu(reap_node
, cpu
) = node
;
909 static void next_reap_node(void)
911 int node
= __get_cpu_var(reap_node
);
913 node
= next_node(node
, node_online_map
);
914 if (unlikely(node
>= MAX_NUMNODES
))
915 node
= first_node(node_online_map
);
916 __get_cpu_var(reap_node
) = node
;
920 #define init_reap_node(cpu) do { } while (0)
921 #define next_reap_node(void) do { } while (0)
925 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
926 * via the workqueue/eventd.
927 * Add the CPU number into the expiration time to minimize the possibility of
928 * the CPUs getting into lockstep and contending for the global cache chain
931 static void __devinit
start_cpu_timer(int cpu
)
933 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
936 * When this gets called from do_initcalls via cpucache_init(),
937 * init_workqueues() has already run, so keventd will be setup
940 if (keventd_up() && reap_work
->work
.func
== NULL
) {
942 INIT_DELAYED_WORK(reap_work
, cache_reap
);
943 schedule_delayed_work_on(cpu
, reap_work
,
944 __round_jiffies_relative(HZ
, cpu
));
948 static struct array_cache
*alloc_arraycache(int node
, int entries
,
951 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
952 struct array_cache
*nc
= NULL
;
954 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
958 nc
->batchcount
= batchcount
;
960 spin_lock_init(&nc
->lock
);
966 * Transfer objects in one arraycache to another.
967 * Locking must be handled by the caller.
969 * Return the number of entries transferred.
971 static int transfer_objects(struct array_cache
*to
,
972 struct array_cache
*from
, unsigned int max
)
974 /* Figure out how many entries to transfer */
975 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
980 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
991 #define drain_alien_cache(cachep, alien) do { } while (0)
992 #define reap_alien(cachep, l3) do { } while (0)
994 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
996 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
999 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1003 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1008 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1014 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1015 gfp_t flags
, int nodeid
)
1020 #else /* CONFIG_NUMA */
1022 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1023 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1025 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
1027 struct array_cache
**ac_ptr
;
1028 int memsize
= sizeof(void *) * nr_node_ids
;
1033 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1036 if (i
== node
|| !node_online(i
)) {
1040 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
1042 for (i
--; i
<= 0; i
--)
1052 static void free_alien_cache(struct array_cache
**ac_ptr
)
1063 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1064 struct array_cache
*ac
, int node
)
1066 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1069 spin_lock(&rl3
->list_lock
);
1071 * Stuff objects into the remote nodes shared array first.
1072 * That way we could avoid the overhead of putting the objects
1073 * into the free lists and getting them back later.
1076 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1078 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1080 spin_unlock(&rl3
->list_lock
);
1085 * Called from cache_reap() to regularly drain alien caches round robin.
1087 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1089 int node
= __get_cpu_var(reap_node
);
1092 struct array_cache
*ac
= l3
->alien
[node
];
1094 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1095 __drain_alien_cache(cachep
, ac
, node
);
1096 spin_unlock_irq(&ac
->lock
);
1101 static void drain_alien_cache(struct kmem_cache
*cachep
,
1102 struct array_cache
**alien
)
1105 struct array_cache
*ac
;
1106 unsigned long flags
;
1108 for_each_online_node(i
) {
1111 spin_lock_irqsave(&ac
->lock
, flags
);
1112 __drain_alien_cache(cachep
, ac
, i
);
1113 spin_unlock_irqrestore(&ac
->lock
, flags
);
1118 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1120 struct slab
*slabp
= virt_to_slab(objp
);
1121 int nodeid
= slabp
->nodeid
;
1122 struct kmem_list3
*l3
;
1123 struct array_cache
*alien
= NULL
;
1126 node
= numa_node_id();
1129 * Make sure we are not freeing a object from another node to the array
1130 * cache on this cpu.
1132 if (likely(slabp
->nodeid
== node
))
1135 l3
= cachep
->nodelists
[node
];
1136 STATS_INC_NODEFREES(cachep
);
1137 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1138 alien
= l3
->alien
[nodeid
];
1139 spin_lock(&alien
->lock
);
1140 if (unlikely(alien
->avail
== alien
->limit
)) {
1141 STATS_INC_ACOVERFLOW(cachep
);
1142 __drain_alien_cache(cachep
, alien
, nodeid
);
1144 alien
->entry
[alien
->avail
++] = objp
;
1145 spin_unlock(&alien
->lock
);
1147 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1148 free_block(cachep
, &objp
, 1, nodeid
);
1149 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1155 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1156 unsigned long action
, void *hcpu
)
1158 long cpu
= (long)hcpu
;
1159 struct kmem_cache
*cachep
;
1160 struct kmem_list3
*l3
= NULL
;
1161 int node
= cpu_to_node(cpu
);
1162 int memsize
= sizeof(struct kmem_list3
);
1165 case CPU_LOCK_ACQUIRE
:
1166 mutex_lock(&cache_chain_mutex
);
1168 case CPU_UP_PREPARE
:
1169 case CPU_UP_PREPARE_FROZEN
:
1171 * We need to do this right in the beginning since
1172 * alloc_arraycache's are going to use this list.
1173 * kmalloc_node allows us to add the slab to the right
1174 * kmem_list3 and not this cpu's kmem_list3
1177 list_for_each_entry(cachep
, &cache_chain
, next
) {
1179 * Set up the size64 kmemlist for cpu before we can
1180 * begin anything. Make sure some other cpu on this
1181 * node has not already allocated this
1183 if (!cachep
->nodelists
[node
]) {
1184 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1187 kmem_list3_init(l3
);
1188 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1189 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1192 * The l3s don't come and go as CPUs come and
1193 * go. cache_chain_mutex is sufficient
1196 cachep
->nodelists
[node
] = l3
;
1199 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1200 cachep
->nodelists
[node
]->free_limit
=
1201 (1 + nr_cpus_node(node
)) *
1202 cachep
->batchcount
+ cachep
->num
;
1203 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1207 * Now we can go ahead with allocating the shared arrays and
1210 list_for_each_entry(cachep
, &cache_chain
, next
) {
1211 struct array_cache
*nc
;
1212 struct array_cache
*shared
= NULL
;
1213 struct array_cache
**alien
= NULL
;
1215 nc
= alloc_arraycache(node
, cachep
->limit
,
1216 cachep
->batchcount
);
1219 if (cachep
->shared
) {
1220 shared
= alloc_arraycache(node
,
1221 cachep
->shared
* cachep
->batchcount
,
1226 if (use_alien_caches
) {
1227 alien
= alloc_alien_cache(node
, cachep
->limit
);
1231 cachep
->array
[cpu
] = nc
;
1232 l3
= cachep
->nodelists
[node
];
1235 spin_lock_irq(&l3
->list_lock
);
1238 * We are serialised from CPU_DEAD or
1239 * CPU_UP_CANCELLED by the cpucontrol lock
1241 l3
->shared
= shared
;
1250 spin_unlock_irq(&l3
->list_lock
);
1252 free_alien_cache(alien
);
1256 case CPU_ONLINE_FROZEN
:
1257 start_cpu_timer(cpu
);
1259 #ifdef CONFIG_HOTPLUG_CPU
1260 case CPU_DOWN_PREPARE
:
1261 case CPU_DOWN_PREPARE_FROZEN
:
1263 * Shutdown cache reaper. Note that the cache_chain_mutex is
1264 * held so that if cache_reap() is invoked it cannot do
1265 * anything expensive but will only modify reap_work
1266 * and reschedule the timer.
1268 cancel_rearming_delayed_work(&per_cpu(reap_work
, cpu
));
1269 /* Now the cache_reaper is guaranteed to be not running. */
1270 per_cpu(reap_work
, cpu
).work
.func
= NULL
;
1272 case CPU_DOWN_FAILED
:
1273 case CPU_DOWN_FAILED_FROZEN
:
1274 start_cpu_timer(cpu
);
1277 case CPU_DEAD_FROZEN
:
1279 * Even if all the cpus of a node are down, we don't free the
1280 * kmem_list3 of any cache. This to avoid a race between
1281 * cpu_down, and a kmalloc allocation from another cpu for
1282 * memory from the node of the cpu going down. The list3
1283 * structure is usually allocated from kmem_cache_create() and
1284 * gets destroyed at kmem_cache_destroy().
1288 case CPU_UP_CANCELED
:
1289 case CPU_UP_CANCELED_FROZEN
:
1290 list_for_each_entry(cachep
, &cache_chain
, next
) {
1291 struct array_cache
*nc
;
1292 struct array_cache
*shared
;
1293 struct array_cache
**alien
;
1296 mask
= node_to_cpumask(node
);
1297 /* cpu is dead; no one can alloc from it. */
1298 nc
= cachep
->array
[cpu
];
1299 cachep
->array
[cpu
] = NULL
;
1300 l3
= cachep
->nodelists
[node
];
1303 goto free_array_cache
;
1305 spin_lock_irq(&l3
->list_lock
);
1307 /* Free limit for this kmem_list3 */
1308 l3
->free_limit
-= cachep
->batchcount
;
1310 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1312 if (!cpus_empty(mask
)) {
1313 spin_unlock_irq(&l3
->list_lock
);
1314 goto free_array_cache
;
1317 shared
= l3
->shared
;
1319 free_block(cachep
, shared
->entry
,
1320 shared
->avail
, node
);
1327 spin_unlock_irq(&l3
->list_lock
);
1331 drain_alien_cache(cachep
, alien
);
1332 free_alien_cache(alien
);
1338 * In the previous loop, all the objects were freed to
1339 * the respective cache's slabs, now we can go ahead and
1340 * shrink each nodelist to its limit.
1342 list_for_each_entry(cachep
, &cache_chain
, next
) {
1343 l3
= cachep
->nodelists
[node
];
1346 drain_freelist(cachep
, l3
, l3
->free_objects
);
1349 case CPU_LOCK_RELEASE
:
1350 mutex_unlock(&cache_chain_mutex
);
1358 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1359 &cpuup_callback
, NULL
, 0
1363 * swap the static kmem_list3 with kmalloced memory
1365 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1368 struct kmem_list3
*ptr
;
1370 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1373 local_irq_disable();
1374 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1376 * Do not assume that spinlocks can be initialized via memcpy:
1378 spin_lock_init(&ptr
->list_lock
);
1380 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1381 cachep
->nodelists
[nodeid
] = ptr
;
1386 * Initialisation. Called after the page allocator have been initialised and
1387 * before smp_init().
1389 void __init
kmem_cache_init(void)
1392 struct cache_sizes
*sizes
;
1393 struct cache_names
*names
;
1398 if (num_possible_nodes() == 1)
1399 use_alien_caches
= 0;
1401 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1402 kmem_list3_init(&initkmem_list3
[i
]);
1403 if (i
< MAX_NUMNODES
)
1404 cache_cache
.nodelists
[i
] = NULL
;
1408 * Fragmentation resistance on low memory - only use bigger
1409 * page orders on machines with more than 32MB of memory.
1411 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1412 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1414 /* Bootstrap is tricky, because several objects are allocated
1415 * from caches that do not exist yet:
1416 * 1) initialize the cache_cache cache: it contains the struct
1417 * kmem_cache structures of all caches, except cache_cache itself:
1418 * cache_cache is statically allocated.
1419 * Initially an __init data area is used for the head array and the
1420 * kmem_list3 structures, it's replaced with a kmalloc allocated
1421 * array at the end of the bootstrap.
1422 * 2) Create the first kmalloc cache.
1423 * The struct kmem_cache for the new cache is allocated normally.
1424 * An __init data area is used for the head array.
1425 * 3) Create the remaining kmalloc caches, with minimally sized
1427 * 4) Replace the __init data head arrays for cache_cache and the first
1428 * kmalloc cache with kmalloc allocated arrays.
1429 * 5) Replace the __init data for kmem_list3 for cache_cache and
1430 * the other cache's with kmalloc allocated memory.
1431 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1434 node
= numa_node_id();
1436 /* 1) create the cache_cache */
1437 INIT_LIST_HEAD(&cache_chain
);
1438 list_add(&cache_cache
.next
, &cache_chain
);
1439 cache_cache
.colour_off
= cache_line_size();
1440 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1441 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
];
1444 * struct kmem_cache size depends on nr_node_ids, which
1445 * can be less than MAX_NUMNODES.
1447 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, nodelists
) +
1448 nr_node_ids
* sizeof(struct kmem_list3
*);
1450 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1452 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1454 cache_cache
.reciprocal_buffer_size
=
1455 reciprocal_value(cache_cache
.buffer_size
);
1457 for (order
= 0; order
< MAX_ORDER
; order
++) {
1458 cache_estimate(order
, cache_cache
.buffer_size
,
1459 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1460 if (cache_cache
.num
)
1463 BUG_ON(!cache_cache
.num
);
1464 cache_cache
.gfporder
= order
;
1465 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1466 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1467 sizeof(struct slab
), cache_line_size());
1469 /* 2+3) create the kmalloc caches */
1470 sizes
= malloc_sizes
;
1471 names
= cache_names
;
1474 * Initialize the caches that provide memory for the array cache and the
1475 * kmem_list3 structures first. Without this, further allocations will
1479 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1480 sizes
[INDEX_AC
].cs_size
,
1481 ARCH_KMALLOC_MINALIGN
,
1482 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1485 if (INDEX_AC
!= INDEX_L3
) {
1486 sizes
[INDEX_L3
].cs_cachep
=
1487 kmem_cache_create(names
[INDEX_L3
].name
,
1488 sizes
[INDEX_L3
].cs_size
,
1489 ARCH_KMALLOC_MINALIGN
,
1490 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1494 slab_early_init
= 0;
1496 while (sizes
->cs_size
!= ULONG_MAX
) {
1498 * For performance, all the general caches are L1 aligned.
1499 * This should be particularly beneficial on SMP boxes, as it
1500 * eliminates "false sharing".
1501 * Note for systems short on memory removing the alignment will
1502 * allow tighter packing of the smaller caches.
1504 if (!sizes
->cs_cachep
) {
1505 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1507 ARCH_KMALLOC_MINALIGN
,
1508 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1511 #ifdef CONFIG_ZONE_DMA
1512 sizes
->cs_dmacachep
= kmem_cache_create(
1515 ARCH_KMALLOC_MINALIGN
,
1516 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1523 /* 4) Replace the bootstrap head arrays */
1525 struct array_cache
*ptr
;
1527 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1529 local_irq_disable();
1530 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1531 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1532 sizeof(struct arraycache_init
));
1534 * Do not assume that spinlocks can be initialized via memcpy:
1536 spin_lock_init(&ptr
->lock
);
1538 cache_cache
.array
[smp_processor_id()] = ptr
;
1541 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1543 local_irq_disable();
1544 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1545 != &initarray_generic
.cache
);
1546 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1547 sizeof(struct arraycache_init
));
1549 * Do not assume that spinlocks can be initialized via memcpy:
1551 spin_lock_init(&ptr
->lock
);
1553 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1557 /* 5) Replace the bootstrap kmem_list3's */
1561 /* Replace the static kmem_list3 structures for the boot cpu */
1562 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
], node
);
1564 for_each_online_node(nid
) {
1565 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1566 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1568 if (INDEX_AC
!= INDEX_L3
) {
1569 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1570 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1575 /* 6) resize the head arrays to their final sizes */
1577 struct kmem_cache
*cachep
;
1578 mutex_lock(&cache_chain_mutex
);
1579 list_for_each_entry(cachep
, &cache_chain
, next
)
1580 if (enable_cpucache(cachep
))
1582 mutex_unlock(&cache_chain_mutex
);
1585 /* Annotate slab for lockdep -- annotate the malloc caches */
1590 g_cpucache_up
= FULL
;
1593 * Register a cpu startup notifier callback that initializes
1594 * cpu_cache_get for all new cpus
1596 register_cpu_notifier(&cpucache_notifier
);
1599 * The reap timers are started later, with a module init call: That part
1600 * of the kernel is not yet operational.
1604 static int __init
cpucache_init(void)
1609 * Register the timers that return unneeded pages to the page allocator
1611 for_each_online_cpu(cpu
)
1612 start_cpu_timer(cpu
);
1615 __initcall(cpucache_init
);
1618 * Interface to system's page allocator. No need to hold the cache-lock.
1620 * If we requested dmaable memory, we will get it. Even if we
1621 * did not request dmaable memory, we might get it, but that
1622 * would be relatively rare and ignorable.
1624 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1632 * Nommu uses slab's for process anonymous memory allocations, and thus
1633 * requires __GFP_COMP to properly refcount higher order allocations
1635 flags
|= __GFP_COMP
;
1638 flags
|= cachep
->gfpflags
;
1640 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1644 nr_pages
= (1 << cachep
->gfporder
);
1645 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1646 add_zone_page_state(page_zone(page
),
1647 NR_SLAB_RECLAIMABLE
, nr_pages
);
1649 add_zone_page_state(page_zone(page
),
1650 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1651 for (i
= 0; i
< nr_pages
; i
++)
1652 __SetPageSlab(page
+ i
);
1653 return page_address(page
);
1657 * Interface to system's page release.
1659 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1661 unsigned long i
= (1 << cachep
->gfporder
);
1662 struct page
*page
= virt_to_page(addr
);
1663 const unsigned long nr_freed
= i
;
1665 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1666 sub_zone_page_state(page_zone(page
),
1667 NR_SLAB_RECLAIMABLE
, nr_freed
);
1669 sub_zone_page_state(page_zone(page
),
1670 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1672 BUG_ON(!PageSlab(page
));
1673 __ClearPageSlab(page
);
1676 if (current
->reclaim_state
)
1677 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1678 free_pages((unsigned long)addr
, cachep
->gfporder
);
1681 static void kmem_rcu_free(struct rcu_head
*head
)
1683 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1684 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1686 kmem_freepages(cachep
, slab_rcu
->addr
);
1687 if (OFF_SLAB(cachep
))
1688 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1693 #ifdef CONFIG_DEBUG_PAGEALLOC
1694 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1695 unsigned long caller
)
1697 int size
= obj_size(cachep
);
1699 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1701 if (size
< 5 * sizeof(unsigned long))
1704 *addr
++ = 0x12345678;
1706 *addr
++ = smp_processor_id();
1707 size
-= 3 * sizeof(unsigned long);
1709 unsigned long *sptr
= &caller
;
1710 unsigned long svalue
;
1712 while (!kstack_end(sptr
)) {
1714 if (kernel_text_address(svalue
)) {
1716 size
-= sizeof(unsigned long);
1717 if (size
<= sizeof(unsigned long))
1723 *addr
++ = 0x87654321;
1727 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1729 int size
= obj_size(cachep
);
1730 addr
= &((char *)addr
)[obj_offset(cachep
)];
1732 memset(addr
, val
, size
);
1733 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1736 static void dump_line(char *data
, int offset
, int limit
)
1739 unsigned char error
= 0;
1742 printk(KERN_ERR
"%03x:", offset
);
1743 for (i
= 0; i
< limit
; i
++) {
1744 if (data
[offset
+ i
] != POISON_FREE
) {
1745 error
= data
[offset
+ i
];
1748 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1752 if (bad_count
== 1) {
1753 error
^= POISON_FREE
;
1754 if (!(error
& (error
- 1))) {
1755 printk(KERN_ERR
"Single bit error detected. Probably "
1758 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1761 printk(KERN_ERR
"Run a memory test tool.\n");
1770 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1775 if (cachep
->flags
& SLAB_RED_ZONE
) {
1776 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1777 *dbg_redzone1(cachep
, objp
),
1778 *dbg_redzone2(cachep
, objp
));
1781 if (cachep
->flags
& SLAB_STORE_USER
) {
1782 printk(KERN_ERR
"Last user: [<%p>]",
1783 *dbg_userword(cachep
, objp
));
1784 print_symbol("(%s)",
1785 (unsigned long)*dbg_userword(cachep
, objp
));
1788 realobj
= (char *)objp
+ obj_offset(cachep
);
1789 size
= obj_size(cachep
);
1790 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1793 if (i
+ limit
> size
)
1795 dump_line(realobj
, i
, limit
);
1799 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1805 realobj
= (char *)objp
+ obj_offset(cachep
);
1806 size
= obj_size(cachep
);
1808 for (i
= 0; i
< size
; i
++) {
1809 char exp
= POISON_FREE
;
1812 if (realobj
[i
] != exp
) {
1818 "Slab corruption: %s start=%p, len=%d\n",
1819 cachep
->name
, realobj
, size
);
1820 print_objinfo(cachep
, objp
, 0);
1822 /* Hexdump the affected line */
1825 if (i
+ limit
> size
)
1827 dump_line(realobj
, i
, limit
);
1830 /* Limit to 5 lines */
1836 /* Print some data about the neighboring objects, if they
1839 struct slab
*slabp
= virt_to_slab(objp
);
1842 objnr
= obj_to_index(cachep
, slabp
, objp
);
1844 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1845 realobj
= (char *)objp
+ obj_offset(cachep
);
1846 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1848 print_objinfo(cachep
, objp
, 2);
1850 if (objnr
+ 1 < cachep
->num
) {
1851 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1852 realobj
= (char *)objp
+ obj_offset(cachep
);
1853 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1855 print_objinfo(cachep
, objp
, 2);
1863 * slab_destroy_objs - destroy a slab and its objects
1864 * @cachep: cache pointer being destroyed
1865 * @slabp: slab pointer being destroyed
1867 * Call the registered destructor for each object in a slab that is being
1870 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1873 for (i
= 0; i
< cachep
->num
; i
++) {
1874 void *objp
= index_to_obj(cachep
, slabp
, i
);
1876 if (cachep
->flags
& SLAB_POISON
) {
1877 #ifdef CONFIG_DEBUG_PAGEALLOC
1878 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1880 kernel_map_pages(virt_to_page(objp
),
1881 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1883 check_poison_obj(cachep
, objp
);
1885 check_poison_obj(cachep
, objp
);
1888 if (cachep
->flags
& SLAB_RED_ZONE
) {
1889 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1890 slab_error(cachep
, "start of a freed object "
1892 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1893 slab_error(cachep
, "end of a freed object "
1899 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1905 * slab_destroy - destroy and release all objects in a slab
1906 * @cachep: cache pointer being destroyed
1907 * @slabp: slab pointer being destroyed
1909 * Destroy all the objs in a slab, and release the mem back to the system.
1910 * Before calling the slab must have been unlinked from the cache. The
1911 * cache-lock is not held/needed.
1913 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1915 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1917 slab_destroy_objs(cachep
, slabp
);
1918 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1919 struct slab_rcu
*slab_rcu
;
1921 slab_rcu
= (struct slab_rcu
*)slabp
;
1922 slab_rcu
->cachep
= cachep
;
1923 slab_rcu
->addr
= addr
;
1924 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1926 kmem_freepages(cachep
, addr
);
1927 if (OFF_SLAB(cachep
))
1928 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1933 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1934 * size of kmem_list3.
1936 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1940 for_each_online_node(node
) {
1941 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1942 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1944 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1948 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1951 struct kmem_list3
*l3
;
1953 for_each_online_cpu(i
)
1954 kfree(cachep
->array
[i
]);
1956 /* NUMA: free the list3 structures */
1957 for_each_online_node(i
) {
1958 l3
= cachep
->nodelists
[i
];
1961 free_alien_cache(l3
->alien
);
1965 kmem_cache_free(&cache_cache
, cachep
);
1970 * calculate_slab_order - calculate size (page order) of slabs
1971 * @cachep: pointer to the cache that is being created
1972 * @size: size of objects to be created in this cache.
1973 * @align: required alignment for the objects.
1974 * @flags: slab allocation flags
1976 * Also calculates the number of objects per slab.
1978 * This could be made much more intelligent. For now, try to avoid using
1979 * high order pages for slabs. When the gfp() functions are more friendly
1980 * towards high-order requests, this should be changed.
1982 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1983 size_t size
, size_t align
, unsigned long flags
)
1985 unsigned long offslab_limit
;
1986 size_t left_over
= 0;
1989 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1993 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1997 if (flags
& CFLGS_OFF_SLAB
) {
1999 * Max number of objs-per-slab for caches which
2000 * use off-slab slabs. Needed to avoid a possible
2001 * looping condition in cache_grow().
2003 offslab_limit
= size
- sizeof(struct slab
);
2004 offslab_limit
/= sizeof(kmem_bufctl_t
);
2006 if (num
> offslab_limit
)
2010 /* Found something acceptable - save it away */
2012 cachep
->gfporder
= gfporder
;
2013 left_over
= remainder
;
2016 * A VFS-reclaimable slab tends to have most allocations
2017 * as GFP_NOFS and we really don't want to have to be allocating
2018 * higher-order pages when we are unable to shrink dcache.
2020 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2024 * Large number of objects is good, but very large slabs are
2025 * currently bad for the gfp()s.
2027 if (gfporder
>= slab_break_gfp_order
)
2031 * Acceptable internal fragmentation?
2033 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2039 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
)
2041 if (g_cpucache_up
== FULL
)
2042 return enable_cpucache(cachep
);
2044 if (g_cpucache_up
== NONE
) {
2046 * Note: the first kmem_cache_create must create the cache
2047 * that's used by kmalloc(24), otherwise the creation of
2048 * further caches will BUG().
2050 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2053 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2054 * the first cache, then we need to set up all its list3s,
2055 * otherwise the creation of further caches will BUG().
2057 set_up_list3s(cachep
, SIZE_AC
);
2058 if (INDEX_AC
== INDEX_L3
)
2059 g_cpucache_up
= PARTIAL_L3
;
2061 g_cpucache_up
= PARTIAL_AC
;
2063 cachep
->array
[smp_processor_id()] =
2064 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
2066 if (g_cpucache_up
== PARTIAL_AC
) {
2067 set_up_list3s(cachep
, SIZE_L3
);
2068 g_cpucache_up
= PARTIAL_L3
;
2071 for_each_online_node(node
) {
2072 cachep
->nodelists
[node
] =
2073 kmalloc_node(sizeof(struct kmem_list3
),
2075 BUG_ON(!cachep
->nodelists
[node
]);
2076 kmem_list3_init(cachep
->nodelists
[node
]);
2080 cachep
->nodelists
[numa_node_id()]->next_reap
=
2081 jiffies
+ REAPTIMEOUT_LIST3
+
2082 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2084 cpu_cache_get(cachep
)->avail
= 0;
2085 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2086 cpu_cache_get(cachep
)->batchcount
= 1;
2087 cpu_cache_get(cachep
)->touched
= 0;
2088 cachep
->batchcount
= 1;
2089 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2094 * kmem_cache_create - Create a cache.
2095 * @name: A string which is used in /proc/slabinfo to identify this cache.
2096 * @size: The size of objects to be created in this cache.
2097 * @align: The required alignment for the objects.
2098 * @flags: SLAB flags
2099 * @ctor: A constructor for the objects.
2100 * @dtor: A destructor for the objects (not implemented anymore).
2102 * Returns a ptr to the cache on success, NULL on failure.
2103 * Cannot be called within a int, but can be interrupted.
2104 * The @ctor is run when new pages are allocated by the cache
2105 * and the @dtor is run before the pages are handed back.
2107 * @name must be valid until the cache is destroyed. This implies that
2108 * the module calling this has to destroy the cache before getting unloaded.
2112 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2113 * to catch references to uninitialised memory.
2115 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2116 * for buffer overruns.
2118 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2119 * cacheline. This can be beneficial if you're counting cycles as closely
2123 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2124 unsigned long flags
,
2125 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
2126 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
2128 size_t left_over
, slab_size
, ralign
;
2129 struct kmem_cache
*cachep
= NULL
, *pc
;
2132 * Sanity checks... these are all serious usage bugs.
2134 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2135 size
> KMALLOC_MAX_SIZE
|| dtor
) {
2136 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
2142 * We use cache_chain_mutex to ensure a consistent view of
2143 * cpu_online_map as well. Please see cpuup_callback
2145 mutex_lock(&cache_chain_mutex
);
2147 list_for_each_entry(pc
, &cache_chain
, next
) {
2152 * This happens when the module gets unloaded and doesn't
2153 * destroy its slab cache and no-one else reuses the vmalloc
2154 * area of the module. Print a warning.
2156 res
= probe_kernel_address(pc
->name
, tmp
);
2159 "SLAB: cache with size %d has lost its name\n",
2164 if (!strcmp(pc
->name
, name
)) {
2166 "kmem_cache_create: duplicate cache %s\n", name
);
2173 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2176 * Enable redzoning and last user accounting, except for caches with
2177 * large objects, if the increased size would increase the object size
2178 * above the next power of two: caches with object sizes just above a
2179 * power of two have a significant amount of internal fragmentation.
2181 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2182 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2183 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2184 flags
|= SLAB_POISON
;
2186 if (flags
& SLAB_DESTROY_BY_RCU
)
2187 BUG_ON(flags
& SLAB_POISON
);
2190 * Always checks flags, a caller might be expecting debug support which
2193 BUG_ON(flags
& ~CREATE_MASK
);
2196 * Check that size is in terms of words. This is needed to avoid
2197 * unaligned accesses for some archs when redzoning is used, and makes
2198 * sure any on-slab bufctl's are also correctly aligned.
2200 if (size
& (BYTES_PER_WORD
- 1)) {
2201 size
+= (BYTES_PER_WORD
- 1);
2202 size
&= ~(BYTES_PER_WORD
- 1);
2205 /* calculate the final buffer alignment: */
2207 /* 1) arch recommendation: can be overridden for debug */
2208 if (flags
& SLAB_HWCACHE_ALIGN
) {
2210 * Default alignment: as specified by the arch code. Except if
2211 * an object is really small, then squeeze multiple objects into
2214 ralign
= cache_line_size();
2215 while (size
<= ralign
/ 2)
2218 ralign
= BYTES_PER_WORD
;
2222 * Redzoning and user store require word alignment. Note this will be
2223 * overridden by architecture or caller mandated alignment if either
2224 * is greater than BYTES_PER_WORD.
2226 if (flags
& SLAB_RED_ZONE
|| flags
& SLAB_STORE_USER
)
2227 ralign
= __alignof__(unsigned long long);
2229 /* 2) arch mandated alignment */
2230 if (ralign
< ARCH_SLAB_MINALIGN
) {
2231 ralign
= ARCH_SLAB_MINALIGN
;
2233 /* 3) caller mandated alignment */
2234 if (ralign
< align
) {
2237 /* disable debug if necessary */
2238 if (ralign
> __alignof__(unsigned long long))
2239 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2245 /* Get cache's description obj. */
2246 cachep
= kmem_cache_zalloc(&cache_cache
, GFP_KERNEL
);
2251 cachep
->obj_size
= size
;
2254 * Both debugging options require word-alignment which is calculated
2257 if (flags
& SLAB_RED_ZONE
) {
2258 /* add space for red zone words */
2259 cachep
->obj_offset
+= sizeof(unsigned long long);
2260 size
+= 2 * sizeof(unsigned long long);
2262 if (flags
& SLAB_STORE_USER
) {
2263 /* user store requires one word storage behind the end of
2266 size
+= BYTES_PER_WORD
;
2268 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2269 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2270 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2271 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2278 * Determine if the slab management is 'on' or 'off' slab.
2279 * (bootstrapping cannot cope with offslab caches so don't do
2282 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2284 * Size is large, assume best to place the slab management obj
2285 * off-slab (should allow better packing of objs).
2287 flags
|= CFLGS_OFF_SLAB
;
2289 size
= ALIGN(size
, align
);
2291 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2295 "kmem_cache_create: couldn't create cache %s.\n", name
);
2296 kmem_cache_free(&cache_cache
, cachep
);
2300 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2301 + sizeof(struct slab
), align
);
2304 * If the slab has been placed off-slab, and we have enough space then
2305 * move it on-slab. This is at the expense of any extra colouring.
2307 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2308 flags
&= ~CFLGS_OFF_SLAB
;
2309 left_over
-= slab_size
;
2312 if (flags
& CFLGS_OFF_SLAB
) {
2313 /* really off slab. No need for manual alignment */
2315 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2318 cachep
->colour_off
= cache_line_size();
2319 /* Offset must be a multiple of the alignment. */
2320 if (cachep
->colour_off
< align
)
2321 cachep
->colour_off
= align
;
2322 cachep
->colour
= left_over
/ cachep
->colour_off
;
2323 cachep
->slab_size
= slab_size
;
2324 cachep
->flags
= flags
;
2325 cachep
->gfpflags
= 0;
2326 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2327 cachep
->gfpflags
|= GFP_DMA
;
2328 cachep
->buffer_size
= size
;
2329 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2331 if (flags
& CFLGS_OFF_SLAB
) {
2332 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2334 * This is a possibility for one of the malloc_sizes caches.
2335 * But since we go off slab only for object size greater than
2336 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2337 * this should not happen at all.
2338 * But leave a BUG_ON for some lucky dude.
2340 BUG_ON(!cachep
->slabp_cache
);
2342 cachep
->ctor
= ctor
;
2343 cachep
->name
= name
;
2345 if (setup_cpu_cache(cachep
)) {
2346 __kmem_cache_destroy(cachep
);
2351 /* cache setup completed, link it into the list */
2352 list_add(&cachep
->next
, &cache_chain
);
2354 if (!cachep
&& (flags
& SLAB_PANIC
))
2355 panic("kmem_cache_create(): failed to create slab `%s'\n",
2357 mutex_unlock(&cache_chain_mutex
);
2360 EXPORT_SYMBOL(kmem_cache_create
);
2363 static void check_irq_off(void)
2365 BUG_ON(!irqs_disabled());
2368 static void check_irq_on(void)
2370 BUG_ON(irqs_disabled());
2373 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2377 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2381 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2385 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2390 #define check_irq_off() do { } while(0)
2391 #define check_irq_on() do { } while(0)
2392 #define check_spinlock_acquired(x) do { } while(0)
2393 #define check_spinlock_acquired_node(x, y) do { } while(0)
2396 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2397 struct array_cache
*ac
,
2398 int force
, int node
);
2400 static void do_drain(void *arg
)
2402 struct kmem_cache
*cachep
= arg
;
2403 struct array_cache
*ac
;
2404 int node
= numa_node_id();
2407 ac
= cpu_cache_get(cachep
);
2408 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2409 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2410 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2414 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2416 struct kmem_list3
*l3
;
2419 on_each_cpu(do_drain
, cachep
, 1, 1);
2421 for_each_online_node(node
) {
2422 l3
= cachep
->nodelists
[node
];
2423 if (l3
&& l3
->alien
)
2424 drain_alien_cache(cachep
, l3
->alien
);
2427 for_each_online_node(node
) {
2428 l3
= cachep
->nodelists
[node
];
2430 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2435 * Remove slabs from the list of free slabs.
2436 * Specify the number of slabs to drain in tofree.
2438 * Returns the actual number of slabs released.
2440 static int drain_freelist(struct kmem_cache
*cache
,
2441 struct kmem_list3
*l3
, int tofree
)
2443 struct list_head
*p
;
2448 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2450 spin_lock_irq(&l3
->list_lock
);
2451 p
= l3
->slabs_free
.prev
;
2452 if (p
== &l3
->slabs_free
) {
2453 spin_unlock_irq(&l3
->list_lock
);
2457 slabp
= list_entry(p
, struct slab
, list
);
2459 BUG_ON(slabp
->inuse
);
2461 list_del(&slabp
->list
);
2463 * Safe to drop the lock. The slab is no longer linked
2466 l3
->free_objects
-= cache
->num
;
2467 spin_unlock_irq(&l3
->list_lock
);
2468 slab_destroy(cache
, slabp
);
2475 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2476 static int __cache_shrink(struct kmem_cache
*cachep
)
2479 struct kmem_list3
*l3
;
2481 drain_cpu_caches(cachep
);
2484 for_each_online_node(i
) {
2485 l3
= cachep
->nodelists
[i
];
2489 drain_freelist(cachep
, l3
, l3
->free_objects
);
2491 ret
+= !list_empty(&l3
->slabs_full
) ||
2492 !list_empty(&l3
->slabs_partial
);
2494 return (ret
? 1 : 0);
2498 * kmem_cache_shrink - Shrink a cache.
2499 * @cachep: The cache to shrink.
2501 * Releases as many slabs as possible for a cache.
2502 * To help debugging, a zero exit status indicates all slabs were released.
2504 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2507 BUG_ON(!cachep
|| in_interrupt());
2509 mutex_lock(&cache_chain_mutex
);
2510 ret
= __cache_shrink(cachep
);
2511 mutex_unlock(&cache_chain_mutex
);
2514 EXPORT_SYMBOL(kmem_cache_shrink
);
2517 * kmem_cache_destroy - delete a cache
2518 * @cachep: the cache to destroy
2520 * Remove a &struct kmem_cache object from the slab cache.
2522 * It is expected this function will be called by a module when it is
2523 * unloaded. This will remove the cache completely, and avoid a duplicate
2524 * cache being allocated each time a module is loaded and unloaded, if the
2525 * module doesn't have persistent in-kernel storage across loads and unloads.
2527 * The cache must be empty before calling this function.
2529 * The caller must guarantee that noone will allocate memory from the cache
2530 * during the kmem_cache_destroy().
2532 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2534 BUG_ON(!cachep
|| in_interrupt());
2536 /* Find the cache in the chain of caches. */
2537 mutex_lock(&cache_chain_mutex
);
2539 * the chain is never empty, cache_cache is never destroyed
2541 list_del(&cachep
->next
);
2542 if (__cache_shrink(cachep
)) {
2543 slab_error(cachep
, "Can't free all objects");
2544 list_add(&cachep
->next
, &cache_chain
);
2545 mutex_unlock(&cache_chain_mutex
);
2549 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2552 __kmem_cache_destroy(cachep
);
2553 mutex_unlock(&cache_chain_mutex
);
2555 EXPORT_SYMBOL(kmem_cache_destroy
);
2558 * Get the memory for a slab management obj.
2559 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2560 * always come from malloc_sizes caches. The slab descriptor cannot
2561 * come from the same cache which is getting created because,
2562 * when we are searching for an appropriate cache for these
2563 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2564 * If we are creating a malloc_sizes cache here it would not be visible to
2565 * kmem_find_general_cachep till the initialization is complete.
2566 * Hence we cannot have slabp_cache same as the original cache.
2568 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2569 int colour_off
, gfp_t local_flags
,
2574 if (OFF_SLAB(cachep
)) {
2575 /* Slab management obj is off-slab. */
2576 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2577 local_flags
& ~GFP_THISNODE
, nodeid
);
2581 slabp
= objp
+ colour_off
;
2582 colour_off
+= cachep
->slab_size
;
2585 slabp
->colouroff
= colour_off
;
2586 slabp
->s_mem
= objp
+ colour_off
;
2587 slabp
->nodeid
= nodeid
;
2591 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2593 return (kmem_bufctl_t
*) (slabp
+ 1);
2596 static void cache_init_objs(struct kmem_cache
*cachep
,
2601 for (i
= 0; i
< cachep
->num
; i
++) {
2602 void *objp
= index_to_obj(cachep
, slabp
, i
);
2604 /* need to poison the objs? */
2605 if (cachep
->flags
& SLAB_POISON
)
2606 poison_obj(cachep
, objp
, POISON_FREE
);
2607 if (cachep
->flags
& SLAB_STORE_USER
)
2608 *dbg_userword(cachep
, objp
) = NULL
;
2610 if (cachep
->flags
& SLAB_RED_ZONE
) {
2611 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2612 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2615 * Constructors are not allowed to allocate memory from the same
2616 * cache which they are a constructor for. Otherwise, deadlock.
2617 * They must also be threaded.
2619 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2620 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2623 if (cachep
->flags
& SLAB_RED_ZONE
) {
2624 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2625 slab_error(cachep
, "constructor overwrote the"
2626 " end of an object");
2627 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2628 slab_error(cachep
, "constructor overwrote the"
2629 " start of an object");
2631 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2632 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2633 kernel_map_pages(virt_to_page(objp
),
2634 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2637 cachep
->ctor(objp
, cachep
, 0);
2639 slab_bufctl(slabp
)[i
] = i
+ 1;
2641 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2645 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2647 if (CONFIG_ZONE_DMA_FLAG
) {
2648 if (flags
& GFP_DMA
)
2649 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2651 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2655 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2658 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2662 next
= slab_bufctl(slabp
)[slabp
->free
];
2664 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2665 WARN_ON(slabp
->nodeid
!= nodeid
);
2672 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2673 void *objp
, int nodeid
)
2675 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2678 /* Verify that the slab belongs to the intended node */
2679 WARN_ON(slabp
->nodeid
!= nodeid
);
2681 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2682 printk(KERN_ERR
"slab: double free detected in cache "
2683 "'%s', objp %p\n", cachep
->name
, objp
);
2687 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2688 slabp
->free
= objnr
;
2693 * Map pages beginning at addr to the given cache and slab. This is required
2694 * for the slab allocator to be able to lookup the cache and slab of a
2695 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2697 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2703 page
= virt_to_page(addr
);
2706 if (likely(!PageCompound(page
)))
2707 nr_pages
<<= cache
->gfporder
;
2710 page_set_cache(page
, cache
);
2711 page_set_slab(page
, slab
);
2713 } while (--nr_pages
);
2717 * Grow (by 1) the number of slabs within a cache. This is called by
2718 * kmem_cache_alloc() when there are no active objs left in a cache.
2720 static int cache_grow(struct kmem_cache
*cachep
,
2721 gfp_t flags
, int nodeid
, void *objp
)
2726 struct kmem_list3
*l3
;
2729 * Be lazy and only check for valid flags here, keeping it out of the
2730 * critical path in kmem_cache_alloc().
2732 BUG_ON(flags
& ~(GFP_DMA
| GFP_LEVEL_MASK
));
2734 local_flags
= (flags
& GFP_LEVEL_MASK
);
2735 /* Take the l3 list lock to change the colour_next on this node */
2737 l3
= cachep
->nodelists
[nodeid
];
2738 spin_lock(&l3
->list_lock
);
2740 /* Get colour for the slab, and cal the next value. */
2741 offset
= l3
->colour_next
;
2743 if (l3
->colour_next
>= cachep
->colour
)
2744 l3
->colour_next
= 0;
2745 spin_unlock(&l3
->list_lock
);
2747 offset
*= cachep
->colour_off
;
2749 if (local_flags
& __GFP_WAIT
)
2753 * The test for missing atomic flag is performed here, rather than
2754 * the more obvious place, simply to reduce the critical path length
2755 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2756 * will eventually be caught here (where it matters).
2758 kmem_flagcheck(cachep
, flags
);
2761 * Get mem for the objs. Attempt to allocate a physical page from
2765 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2769 /* Get slab management. */
2770 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2771 local_flags
& ~GFP_THISNODE
, nodeid
);
2775 slabp
->nodeid
= nodeid
;
2776 slab_map_pages(cachep
, slabp
, objp
);
2778 cache_init_objs(cachep
, slabp
);
2780 if (local_flags
& __GFP_WAIT
)
2781 local_irq_disable();
2783 spin_lock(&l3
->list_lock
);
2785 /* Make slab active. */
2786 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2787 STATS_INC_GROWN(cachep
);
2788 l3
->free_objects
+= cachep
->num
;
2789 spin_unlock(&l3
->list_lock
);
2792 kmem_freepages(cachep
, objp
);
2794 if (local_flags
& __GFP_WAIT
)
2795 local_irq_disable();
2802 * Perform extra freeing checks:
2803 * - detect bad pointers.
2804 * - POISON/RED_ZONE checking
2806 static void kfree_debugcheck(const void *objp
)
2808 if (!virt_addr_valid(objp
)) {
2809 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2810 (unsigned long)objp
);
2815 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2817 unsigned long long redzone1
, redzone2
;
2819 redzone1
= *dbg_redzone1(cache
, obj
);
2820 redzone2
= *dbg_redzone2(cache
, obj
);
2825 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2828 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2829 slab_error(cache
, "double free detected");
2831 slab_error(cache
, "memory outside object was overwritten");
2833 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2834 obj
, redzone1
, redzone2
);
2837 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2844 objp
-= obj_offset(cachep
);
2845 kfree_debugcheck(objp
);
2846 page
= virt_to_head_page(objp
);
2848 slabp
= page_get_slab(page
);
2850 if (cachep
->flags
& SLAB_RED_ZONE
) {
2851 verify_redzone_free(cachep
, objp
);
2852 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2853 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2855 if (cachep
->flags
& SLAB_STORE_USER
)
2856 *dbg_userword(cachep
, objp
) = caller
;
2858 objnr
= obj_to_index(cachep
, slabp
, objp
);
2860 BUG_ON(objnr
>= cachep
->num
);
2861 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2863 #ifdef CONFIG_DEBUG_SLAB_LEAK
2864 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2866 if (cachep
->flags
& SLAB_POISON
) {
2867 #ifdef CONFIG_DEBUG_PAGEALLOC
2868 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2869 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2870 kernel_map_pages(virt_to_page(objp
),
2871 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2873 poison_obj(cachep
, objp
, POISON_FREE
);
2876 poison_obj(cachep
, objp
, POISON_FREE
);
2882 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2887 /* Check slab's freelist to see if this obj is there. */
2888 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2890 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2893 if (entries
!= cachep
->num
- slabp
->inuse
) {
2895 printk(KERN_ERR
"slab: Internal list corruption detected in "
2896 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2897 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2899 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2902 printk("\n%03x:", i
);
2903 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2910 #define kfree_debugcheck(x) do { } while(0)
2911 #define cache_free_debugcheck(x,objp,z) (objp)
2912 #define check_slabp(x,y) do { } while(0)
2915 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2918 struct kmem_list3
*l3
;
2919 struct array_cache
*ac
;
2922 node
= numa_node_id();
2925 ac
= cpu_cache_get(cachep
);
2927 batchcount
= ac
->batchcount
;
2928 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2930 * If there was little recent activity on this cache, then
2931 * perform only a partial refill. Otherwise we could generate
2934 batchcount
= BATCHREFILL_LIMIT
;
2936 l3
= cachep
->nodelists
[node
];
2938 BUG_ON(ac
->avail
> 0 || !l3
);
2939 spin_lock(&l3
->list_lock
);
2941 /* See if we can refill from the shared array */
2942 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2945 while (batchcount
> 0) {
2946 struct list_head
*entry
;
2948 /* Get slab alloc is to come from. */
2949 entry
= l3
->slabs_partial
.next
;
2950 if (entry
== &l3
->slabs_partial
) {
2951 l3
->free_touched
= 1;
2952 entry
= l3
->slabs_free
.next
;
2953 if (entry
== &l3
->slabs_free
)
2957 slabp
= list_entry(entry
, struct slab
, list
);
2958 check_slabp(cachep
, slabp
);
2959 check_spinlock_acquired(cachep
);
2962 * The slab was either on partial or free list so
2963 * there must be at least one object available for
2966 BUG_ON(slabp
->inuse
< 0 || slabp
->inuse
>= cachep
->num
);
2968 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2969 STATS_INC_ALLOCED(cachep
);
2970 STATS_INC_ACTIVE(cachep
);
2971 STATS_SET_HIGH(cachep
);
2973 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2976 check_slabp(cachep
, slabp
);
2978 /* move slabp to correct slabp list: */
2979 list_del(&slabp
->list
);
2980 if (slabp
->free
== BUFCTL_END
)
2981 list_add(&slabp
->list
, &l3
->slabs_full
);
2983 list_add(&slabp
->list
, &l3
->slabs_partial
);
2987 l3
->free_objects
-= ac
->avail
;
2989 spin_unlock(&l3
->list_lock
);
2991 if (unlikely(!ac
->avail
)) {
2993 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2995 /* cache_grow can reenable interrupts, then ac could change. */
2996 ac
= cpu_cache_get(cachep
);
2997 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3000 if (!ac
->avail
) /* objects refilled by interrupt? */
3004 return ac
->entry
[--ac
->avail
];
3007 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3010 might_sleep_if(flags
& __GFP_WAIT
);
3012 kmem_flagcheck(cachep
, flags
);
3017 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3018 gfp_t flags
, void *objp
, void *caller
)
3022 if (cachep
->flags
& SLAB_POISON
) {
3023 #ifdef CONFIG_DEBUG_PAGEALLOC
3024 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3025 kernel_map_pages(virt_to_page(objp
),
3026 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3028 check_poison_obj(cachep
, objp
);
3030 check_poison_obj(cachep
, objp
);
3032 poison_obj(cachep
, objp
, POISON_INUSE
);
3034 if (cachep
->flags
& SLAB_STORE_USER
)
3035 *dbg_userword(cachep
, objp
) = caller
;
3037 if (cachep
->flags
& SLAB_RED_ZONE
) {
3038 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3039 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3040 slab_error(cachep
, "double free, or memory outside"
3041 " object was overwritten");
3043 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3044 objp
, *dbg_redzone1(cachep
, objp
),
3045 *dbg_redzone2(cachep
, objp
));
3047 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3048 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3050 #ifdef CONFIG_DEBUG_SLAB_LEAK
3055 slabp
= page_get_slab(virt_to_head_page(objp
));
3056 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3057 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3060 objp
+= obj_offset(cachep
);
3061 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3062 cachep
->ctor(objp
, cachep
, 0);
3063 #if ARCH_SLAB_MINALIGN
3064 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3065 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3066 objp
, ARCH_SLAB_MINALIGN
);
3072 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3075 #ifdef CONFIG_FAILSLAB
3077 static struct failslab_attr
{
3079 struct fault_attr attr
;
3081 u32 ignore_gfp_wait
;
3082 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3083 struct dentry
*ignore_gfp_wait_file
;
3087 .attr
= FAULT_ATTR_INITIALIZER
,
3088 .ignore_gfp_wait
= 1,
3091 static int __init
setup_failslab(char *str
)
3093 return setup_fault_attr(&failslab
.attr
, str
);
3095 __setup("failslab=", setup_failslab
);
3097 static int should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3099 if (cachep
== &cache_cache
)
3101 if (flags
& __GFP_NOFAIL
)
3103 if (failslab
.ignore_gfp_wait
&& (flags
& __GFP_WAIT
))
3106 return should_fail(&failslab
.attr
, obj_size(cachep
));
3109 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3111 static int __init
failslab_debugfs(void)
3113 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
3117 err
= init_fault_attr_dentries(&failslab
.attr
, "failslab");
3120 dir
= failslab
.attr
.dentries
.dir
;
3122 failslab
.ignore_gfp_wait_file
=
3123 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3124 &failslab
.ignore_gfp_wait
);
3126 if (!failslab
.ignore_gfp_wait_file
) {
3128 debugfs_remove(failslab
.ignore_gfp_wait_file
);
3129 cleanup_fault_attr_dentries(&failslab
.attr
);
3135 late_initcall(failslab_debugfs
);
3137 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3139 #else /* CONFIG_FAILSLAB */
3141 static inline int should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3146 #endif /* CONFIG_FAILSLAB */
3148 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3151 struct array_cache
*ac
;
3155 ac
= cpu_cache_get(cachep
);
3156 if (likely(ac
->avail
)) {
3157 STATS_INC_ALLOCHIT(cachep
);
3159 objp
= ac
->entry
[--ac
->avail
];
3161 STATS_INC_ALLOCMISS(cachep
);
3162 objp
= cache_alloc_refill(cachep
, flags
);
3169 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3171 * If we are in_interrupt, then process context, including cpusets and
3172 * mempolicy, may not apply and should not be used for allocation policy.
3174 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3176 int nid_alloc
, nid_here
;
3178 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3180 nid_alloc
= nid_here
= numa_node_id();
3181 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3182 nid_alloc
= cpuset_mem_spread_node();
3183 else if (current
->mempolicy
)
3184 nid_alloc
= slab_node(current
->mempolicy
);
3185 if (nid_alloc
!= nid_here
)
3186 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3191 * Fallback function if there was no memory available and no objects on a
3192 * certain node and fall back is permitted. First we scan all the
3193 * available nodelists for available objects. If that fails then we
3194 * perform an allocation without specifying a node. This allows the page
3195 * allocator to do its reclaim / fallback magic. We then insert the
3196 * slab into the proper nodelist and then allocate from it.
3198 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3200 struct zonelist
*zonelist
;
3206 if (flags
& __GFP_THISNODE
)
3209 zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
3210 ->node_zonelists
[gfp_zone(flags
)];
3211 local_flags
= (flags
& GFP_LEVEL_MASK
);
3215 * Look through allowed nodes for objects available
3216 * from existing per node queues.
3218 for (z
= zonelist
->zones
; *z
&& !obj
; z
++) {
3219 nid
= zone_to_nid(*z
);
3221 if (cpuset_zone_allowed_hardwall(*z
, flags
) &&
3222 cache
->nodelists
[nid
] &&
3223 cache
->nodelists
[nid
]->free_objects
)
3224 obj
= ____cache_alloc_node(cache
,
3225 flags
| GFP_THISNODE
, nid
);
3230 * This allocation will be performed within the constraints
3231 * of the current cpuset / memory policy requirements.
3232 * We may trigger various forms of reclaim on the allowed
3233 * set and go into memory reserves if necessary.
3235 if (local_flags
& __GFP_WAIT
)
3237 kmem_flagcheck(cache
, flags
);
3238 obj
= kmem_getpages(cache
, flags
, -1);
3239 if (local_flags
& __GFP_WAIT
)
3240 local_irq_disable();
3243 * Insert into the appropriate per node queues
3245 nid
= page_to_nid(virt_to_page(obj
));
3246 if (cache_grow(cache
, flags
, nid
, obj
)) {
3247 obj
= ____cache_alloc_node(cache
,
3248 flags
| GFP_THISNODE
, nid
);
3251 * Another processor may allocate the
3252 * objects in the slab since we are
3253 * not holding any locks.
3257 /* cache_grow already freed obj */
3266 * A interface to enable slab creation on nodeid
3268 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3271 struct list_head
*entry
;
3273 struct kmem_list3
*l3
;
3277 l3
= cachep
->nodelists
[nodeid
];
3282 spin_lock(&l3
->list_lock
);
3283 entry
= l3
->slabs_partial
.next
;
3284 if (entry
== &l3
->slabs_partial
) {
3285 l3
->free_touched
= 1;
3286 entry
= l3
->slabs_free
.next
;
3287 if (entry
== &l3
->slabs_free
)
3291 slabp
= list_entry(entry
, struct slab
, list
);
3292 check_spinlock_acquired_node(cachep
, nodeid
);
3293 check_slabp(cachep
, slabp
);
3295 STATS_INC_NODEALLOCS(cachep
);
3296 STATS_INC_ACTIVE(cachep
);
3297 STATS_SET_HIGH(cachep
);
3299 BUG_ON(slabp
->inuse
== cachep
->num
);
3301 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3302 check_slabp(cachep
, slabp
);
3304 /* move slabp to correct slabp list: */
3305 list_del(&slabp
->list
);
3307 if (slabp
->free
== BUFCTL_END
)
3308 list_add(&slabp
->list
, &l3
->slabs_full
);
3310 list_add(&slabp
->list
, &l3
->slabs_partial
);
3312 spin_unlock(&l3
->list_lock
);
3316 spin_unlock(&l3
->list_lock
);
3317 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3321 return fallback_alloc(cachep
, flags
);
3328 * kmem_cache_alloc_node - Allocate an object on the specified node
3329 * @cachep: The cache to allocate from.
3330 * @flags: See kmalloc().
3331 * @nodeid: node number of the target node.
3332 * @caller: return address of caller, used for debug information
3334 * Identical to kmem_cache_alloc but it will allocate memory on the given
3335 * node, which can improve the performance for cpu bound structures.
3337 * Fallback to other node is possible if __GFP_THISNODE is not set.
3339 static __always_inline
void *
3340 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3343 unsigned long save_flags
;
3346 if (should_failslab(cachep
, flags
))
3349 cache_alloc_debugcheck_before(cachep
, flags
);
3350 local_irq_save(save_flags
);
3352 if (unlikely(nodeid
== -1))
3353 nodeid
= numa_node_id();
3355 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3356 /* Node not bootstrapped yet */
3357 ptr
= fallback_alloc(cachep
, flags
);
3361 if (nodeid
== numa_node_id()) {
3363 * Use the locally cached objects if possible.
3364 * However ____cache_alloc does not allow fallback
3365 * to other nodes. It may fail while we still have
3366 * objects on other nodes available.
3368 ptr
= ____cache_alloc(cachep
, flags
);
3372 /* ___cache_alloc_node can fall back to other nodes */
3373 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3375 local_irq_restore(save_flags
);
3376 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3381 static __always_inline
void *
3382 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3386 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3387 objp
= alternate_node_alloc(cache
, flags
);
3391 objp
= ____cache_alloc(cache
, flags
);
3394 * We may just have run out of memory on the local node.
3395 * ____cache_alloc_node() knows how to locate memory on other nodes
3398 objp
= ____cache_alloc_node(cache
, flags
, numa_node_id());
3405 static __always_inline
void *
3406 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3408 return ____cache_alloc(cachep
, flags
);
3411 #endif /* CONFIG_NUMA */
3413 static __always_inline
void *
3414 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3416 unsigned long save_flags
;
3419 if (should_failslab(cachep
, flags
))
3422 cache_alloc_debugcheck_before(cachep
, flags
);
3423 local_irq_save(save_flags
);
3424 objp
= __do_cache_alloc(cachep
, flags
);
3425 local_irq_restore(save_flags
);
3426 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
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 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3541 if (cache_free_alien(cachep
, objp
))
3544 if (likely(ac
->avail
< ac
->limit
)) {
3545 STATS_INC_FREEHIT(cachep
);
3546 ac
->entry
[ac
->avail
++] = objp
;
3549 STATS_INC_FREEMISS(cachep
);
3550 cache_flusharray(cachep
, ac
);
3551 ac
->entry
[ac
->avail
++] = objp
;
3556 * kmem_cache_alloc - Allocate an object
3557 * @cachep: The cache to allocate from.
3558 * @flags: See kmalloc().
3560 * Allocate an object from this cache. The flags are only relevant
3561 * if the cache has no available objects.
3563 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3565 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3567 EXPORT_SYMBOL(kmem_cache_alloc
);
3570 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3571 * @cache: The cache to allocate from.
3572 * @flags: See kmalloc().
3574 * Allocate an object from this cache and set the allocated memory to zero.
3575 * The flags are only relevant if the cache has no available objects.
3577 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3579 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3581 memset(ret
, 0, obj_size(cache
));
3584 EXPORT_SYMBOL(kmem_cache_zalloc
);
3587 * kmem_ptr_validate - check if an untrusted pointer might
3589 * @cachep: the cache we're checking against
3590 * @ptr: pointer to validate
3592 * This verifies that the untrusted pointer looks sane:
3593 * it is _not_ a guarantee that the pointer is actually
3594 * part of the slab cache in question, but it at least
3595 * validates that the pointer can be dereferenced and
3596 * looks half-way sane.
3598 * Currently only used for dentry validation.
3600 int kmem_ptr_validate(struct kmem_cache
*cachep
, const void *ptr
)
3602 unsigned long addr
= (unsigned long)ptr
;
3603 unsigned long min_addr
= PAGE_OFFSET
;
3604 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3605 unsigned long size
= cachep
->buffer_size
;
3608 if (unlikely(addr
< min_addr
))
3610 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3612 if (unlikely(addr
& align_mask
))
3614 if (unlikely(!kern_addr_valid(addr
)))
3616 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3618 page
= virt_to_page(ptr
);
3619 if (unlikely(!PageSlab(page
)))
3621 if (unlikely(page_get_cache(page
) != cachep
))
3629 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3631 return __cache_alloc_node(cachep
, flags
, nodeid
,
3632 __builtin_return_address(0));
3634 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3636 static __always_inline
void *
3637 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3639 struct kmem_cache
*cachep
;
3641 cachep
= kmem_find_general_cachep(size
, flags
);
3642 if (unlikely(cachep
== NULL
))
3644 return kmem_cache_alloc_node(cachep
, flags
, node
);
3647 #ifdef CONFIG_DEBUG_SLAB
3648 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3650 return __do_kmalloc_node(size
, flags
, node
,
3651 __builtin_return_address(0));
3653 EXPORT_SYMBOL(__kmalloc_node
);
3655 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3656 int node
, void *caller
)
3658 return __do_kmalloc_node(size
, flags
, node
, caller
);
3660 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3662 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3664 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3666 EXPORT_SYMBOL(__kmalloc_node
);
3667 #endif /* CONFIG_DEBUG_SLAB */
3668 #endif /* CONFIG_NUMA */
3671 * __do_kmalloc - allocate memory
3672 * @size: how many bytes of memory are required.
3673 * @flags: the type of memory to allocate (see kmalloc).
3674 * @caller: function caller for debug tracking of the caller
3676 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3679 struct kmem_cache
*cachep
;
3681 /* If you want to save a few bytes .text space: replace
3683 * Then kmalloc uses the uninlined functions instead of the inline
3686 cachep
= __find_general_cachep(size
, flags
);
3687 if (unlikely(cachep
== NULL
))
3689 return __cache_alloc(cachep
, flags
, caller
);
3693 #ifdef CONFIG_DEBUG_SLAB
3694 void *__kmalloc(size_t size
, gfp_t flags
)
3696 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3698 EXPORT_SYMBOL(__kmalloc
);
3700 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3702 return __do_kmalloc(size
, flags
, caller
);
3704 EXPORT_SYMBOL(__kmalloc_track_caller
);
3707 void *__kmalloc(size_t size
, gfp_t flags
)
3709 return __do_kmalloc(size
, flags
, NULL
);
3711 EXPORT_SYMBOL(__kmalloc
);
3715 * krealloc - reallocate memory. The contents will remain unchanged.
3716 * @p: object to reallocate memory for.
3717 * @new_size: how many bytes of memory are required.
3718 * @flags: the type of memory to allocate.
3720 * The contents of the object pointed to are preserved up to the
3721 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3722 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3723 * %NULL pointer, the object pointed to is freed.
3725 void *krealloc(const void *p
, size_t new_size
, gfp_t flags
)
3727 struct kmem_cache
*cache
, *new_cache
;
3731 return kmalloc_track_caller(new_size
, flags
);
3733 if (unlikely(!new_size
)) {
3738 cache
= virt_to_cache(p
);
3739 new_cache
= __find_general_cachep(new_size
, flags
);
3742 * If new size fits in the current cache, bail out.
3744 if (likely(cache
== new_cache
))
3748 * We are on the slow-path here so do not use __cache_alloc
3749 * because it bloats kernel text.
3751 ret
= kmalloc_track_caller(new_size
, flags
);
3753 memcpy(ret
, p
, min(new_size
, ksize(p
)));
3758 EXPORT_SYMBOL(krealloc
);
3761 * kmem_cache_free - Deallocate an object
3762 * @cachep: The cache the allocation was from.
3763 * @objp: The previously allocated object.
3765 * Free an object which was previously allocated from this
3768 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3770 unsigned long flags
;
3772 BUG_ON(virt_to_cache(objp
) != cachep
);
3774 local_irq_save(flags
);
3775 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3776 __cache_free(cachep
, objp
);
3777 local_irq_restore(flags
);
3779 EXPORT_SYMBOL(kmem_cache_free
);
3782 * kfree - free previously allocated memory
3783 * @objp: pointer returned by kmalloc.
3785 * If @objp is NULL, no operation is performed.
3787 * Don't free memory not originally allocated by kmalloc()
3788 * or you will run into trouble.
3790 void kfree(const void *objp
)
3792 struct kmem_cache
*c
;
3793 unsigned long flags
;
3795 if (unlikely(!objp
))
3797 local_irq_save(flags
);
3798 kfree_debugcheck(objp
);
3799 c
= virt_to_cache(objp
);
3800 debug_check_no_locks_freed(objp
, obj_size(c
));
3801 __cache_free(c
, (void *)objp
);
3802 local_irq_restore(flags
);
3804 EXPORT_SYMBOL(kfree
);
3806 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3808 return obj_size(cachep
);
3810 EXPORT_SYMBOL(kmem_cache_size
);
3812 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3814 return cachep
->name
;
3816 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3819 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3821 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3824 struct kmem_list3
*l3
;
3825 struct array_cache
*new_shared
;
3826 struct array_cache
**new_alien
= NULL
;
3828 for_each_online_node(node
) {
3830 if (use_alien_caches
) {
3831 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3837 if (cachep
->shared
) {
3838 new_shared
= alloc_arraycache(node
,
3839 cachep
->shared
*cachep
->batchcount
,
3842 free_alien_cache(new_alien
);
3847 l3
= cachep
->nodelists
[node
];
3849 struct array_cache
*shared
= l3
->shared
;
3851 spin_lock_irq(&l3
->list_lock
);
3854 free_block(cachep
, shared
->entry
,
3855 shared
->avail
, node
);
3857 l3
->shared
= new_shared
;
3859 l3
->alien
= new_alien
;
3862 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3863 cachep
->batchcount
+ cachep
->num
;
3864 spin_unlock_irq(&l3
->list_lock
);
3866 free_alien_cache(new_alien
);
3869 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3871 free_alien_cache(new_alien
);
3876 kmem_list3_init(l3
);
3877 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3878 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3879 l3
->shared
= new_shared
;
3880 l3
->alien
= new_alien
;
3881 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3882 cachep
->batchcount
+ cachep
->num
;
3883 cachep
->nodelists
[node
] = l3
;
3888 if (!cachep
->next
.next
) {
3889 /* Cache is not active yet. Roll back what we did */
3892 if (cachep
->nodelists
[node
]) {
3893 l3
= cachep
->nodelists
[node
];
3896 free_alien_cache(l3
->alien
);
3898 cachep
->nodelists
[node
] = NULL
;
3906 struct ccupdate_struct
{
3907 struct kmem_cache
*cachep
;
3908 struct array_cache
*new[NR_CPUS
];
3911 static void do_ccupdate_local(void *info
)
3913 struct ccupdate_struct
*new = info
;
3914 struct array_cache
*old
;
3917 old
= cpu_cache_get(new->cachep
);
3919 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3920 new->new[smp_processor_id()] = old
;
3923 /* Always called with the cache_chain_mutex held */
3924 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3925 int batchcount
, int shared
)
3927 struct ccupdate_struct
*new;
3930 new = kzalloc(sizeof(*new), GFP_KERNEL
);
3934 for_each_online_cpu(i
) {
3935 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3938 for (i
--; i
>= 0; i
--)
3944 new->cachep
= cachep
;
3946 on_each_cpu(do_ccupdate_local
, (void *)new, 1, 1);
3949 cachep
->batchcount
= batchcount
;
3950 cachep
->limit
= limit
;
3951 cachep
->shared
= shared
;
3953 for_each_online_cpu(i
) {
3954 struct array_cache
*ccold
= new->new[i
];
3957 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3958 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3959 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3963 return alloc_kmemlist(cachep
);
3966 /* Called with cache_chain_mutex held always */
3967 static int enable_cpucache(struct kmem_cache
*cachep
)
3973 * The head array serves three purposes:
3974 * - create a LIFO ordering, i.e. return objects that are cache-warm
3975 * - reduce the number of spinlock operations.
3976 * - reduce the number of linked list operations on the slab and
3977 * bufctl chains: array operations are cheaper.
3978 * The numbers are guessed, we should auto-tune as described by
3981 if (cachep
->buffer_size
> 131072)
3983 else if (cachep
->buffer_size
> PAGE_SIZE
)
3985 else if (cachep
->buffer_size
> 1024)
3987 else if (cachep
->buffer_size
> 256)
3993 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3994 * allocation behaviour: Most allocs on one cpu, most free operations
3995 * on another cpu. For these cases, an efficient object passing between
3996 * cpus is necessary. This is provided by a shared array. The array
3997 * replaces Bonwick's magazine layer.
3998 * On uniprocessor, it's functionally equivalent (but less efficient)
3999 * to a larger limit. Thus disabled by default.
4002 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4007 * With debugging enabled, large batchcount lead to excessively long
4008 * periods with disabled local interrupts. Limit the batchcount
4013 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
4015 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4016 cachep
->name
, -err
);
4021 * Drain an array if it contains any elements taking the l3 lock only if
4022 * necessary. Note that the l3 listlock also protects the array_cache
4023 * if drain_array() is used on the shared array.
4025 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4026 struct array_cache
*ac
, int force
, int node
)
4030 if (!ac
|| !ac
->avail
)
4032 if (ac
->touched
&& !force
) {
4035 spin_lock_irq(&l3
->list_lock
);
4037 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4038 if (tofree
> ac
->avail
)
4039 tofree
= (ac
->avail
+ 1) / 2;
4040 free_block(cachep
, ac
->entry
, tofree
, node
);
4041 ac
->avail
-= tofree
;
4042 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4043 sizeof(void *) * ac
->avail
);
4045 spin_unlock_irq(&l3
->list_lock
);
4050 * cache_reap - Reclaim memory from caches.
4051 * @w: work descriptor
4053 * Called from workqueue/eventd every few seconds.
4055 * - clear the per-cpu caches for this CPU.
4056 * - return freeable pages to the main free memory pool.
4058 * If we cannot acquire the cache chain mutex then just give up - we'll try
4059 * again on the next iteration.
4061 static void cache_reap(struct work_struct
*w
)
4063 struct kmem_cache
*searchp
;
4064 struct kmem_list3
*l3
;
4065 int node
= numa_node_id();
4066 struct delayed_work
*work
=
4067 container_of(w
, struct delayed_work
, work
);
4069 if (!mutex_trylock(&cache_chain_mutex
))
4070 /* Give up. Setup the next iteration. */
4073 list_for_each_entry(searchp
, &cache_chain
, next
) {
4077 * We only take the l3 lock if absolutely necessary and we
4078 * have established with reasonable certainty that
4079 * we can do some work if the lock was obtained.
4081 l3
= searchp
->nodelists
[node
];
4083 reap_alien(searchp
, l3
);
4085 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4088 * These are racy checks but it does not matter
4089 * if we skip one check or scan twice.
4091 if (time_after(l3
->next_reap
, jiffies
))
4094 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4096 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4098 if (l3
->free_touched
)
4099 l3
->free_touched
= 0;
4103 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4104 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4105 STATS_ADD_REAPED(searchp
, freed
);
4111 mutex_unlock(&cache_chain_mutex
);
4114 /* Set up the next iteration */
4115 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4118 #ifdef CONFIG_PROC_FS
4120 static void print_slabinfo_header(struct seq_file
*m
)
4123 * Output format version, so at least we can change it
4124 * without _too_ many complaints.
4127 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4129 seq_puts(m
, "slabinfo - version: 2.1\n");
4131 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4132 "<objperslab> <pagesperslab>");
4133 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4134 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4136 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4137 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4138 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4143 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4146 struct list_head
*p
;
4148 mutex_lock(&cache_chain_mutex
);
4150 print_slabinfo_header(m
);
4151 p
= cache_chain
.next
;
4154 if (p
== &cache_chain
)
4157 return list_entry(p
, struct kmem_cache
, next
);
4160 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4162 struct kmem_cache
*cachep
= p
;
4164 return cachep
->next
.next
== &cache_chain
?
4165 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
4168 static void s_stop(struct seq_file
*m
, void *p
)
4170 mutex_unlock(&cache_chain_mutex
);
4173 static int s_show(struct seq_file
*m
, void *p
)
4175 struct kmem_cache
*cachep
= p
;
4177 unsigned long active_objs
;
4178 unsigned long num_objs
;
4179 unsigned long active_slabs
= 0;
4180 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4184 struct kmem_list3
*l3
;
4188 for_each_online_node(node
) {
4189 l3
= cachep
->nodelists
[node
];
4194 spin_lock_irq(&l3
->list_lock
);
4196 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4197 if (slabp
->inuse
!= cachep
->num
&& !error
)
4198 error
= "slabs_full accounting error";
4199 active_objs
+= cachep
->num
;
4202 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4203 if (slabp
->inuse
== cachep
->num
&& !error
)
4204 error
= "slabs_partial inuse accounting error";
4205 if (!slabp
->inuse
&& !error
)
4206 error
= "slabs_partial/inuse accounting error";
4207 active_objs
+= slabp
->inuse
;
4210 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4211 if (slabp
->inuse
&& !error
)
4212 error
= "slabs_free/inuse accounting error";
4215 free_objects
+= l3
->free_objects
;
4217 shared_avail
+= l3
->shared
->avail
;
4219 spin_unlock_irq(&l3
->list_lock
);
4221 num_slabs
+= active_slabs
;
4222 num_objs
= num_slabs
* cachep
->num
;
4223 if (num_objs
- active_objs
!= free_objects
&& !error
)
4224 error
= "free_objects accounting error";
4226 name
= cachep
->name
;
4228 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4230 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4231 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4232 cachep
->num
, (1 << cachep
->gfporder
));
4233 seq_printf(m
, " : tunables %4u %4u %4u",
4234 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4235 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4236 active_slabs
, num_slabs
, shared_avail
);
4239 unsigned long high
= cachep
->high_mark
;
4240 unsigned long allocs
= cachep
->num_allocations
;
4241 unsigned long grown
= cachep
->grown
;
4242 unsigned long reaped
= cachep
->reaped
;
4243 unsigned long errors
= cachep
->errors
;
4244 unsigned long max_freeable
= cachep
->max_freeable
;
4245 unsigned long node_allocs
= cachep
->node_allocs
;
4246 unsigned long node_frees
= cachep
->node_frees
;
4247 unsigned long overflows
= cachep
->node_overflow
;
4249 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4250 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4251 reaped
, errors
, max_freeable
, node_allocs
,
4252 node_frees
, overflows
);
4256 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4257 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4258 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4259 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4261 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4262 allochit
, allocmiss
, freehit
, freemiss
);
4270 * slabinfo_op - iterator that generates /proc/slabinfo
4279 * num-pages-per-slab
4280 * + further values on SMP and with statistics enabled
4283 const struct seq_operations slabinfo_op
= {
4290 #define MAX_SLABINFO_WRITE 128
4292 * slabinfo_write - Tuning for the slab allocator
4294 * @buffer: user buffer
4295 * @count: data length
4298 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4299 size_t count
, loff_t
*ppos
)
4301 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4302 int limit
, batchcount
, shared
, res
;
4303 struct kmem_cache
*cachep
;
4305 if (count
> MAX_SLABINFO_WRITE
)
4307 if (copy_from_user(&kbuf
, buffer
, count
))
4309 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4311 tmp
= strchr(kbuf
, ' ');
4316 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4319 /* Find the cache in the chain of caches. */
4320 mutex_lock(&cache_chain_mutex
);
4322 list_for_each_entry(cachep
, &cache_chain
, next
) {
4323 if (!strcmp(cachep
->name
, kbuf
)) {
4324 if (limit
< 1 || batchcount
< 1 ||
4325 batchcount
> limit
|| shared
< 0) {
4328 res
= do_tune_cpucache(cachep
, limit
,
4329 batchcount
, shared
);
4334 mutex_unlock(&cache_chain_mutex
);
4340 #ifdef CONFIG_DEBUG_SLAB_LEAK
4342 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4345 struct list_head
*p
;
4347 mutex_lock(&cache_chain_mutex
);
4348 p
= cache_chain
.next
;
4351 if (p
== &cache_chain
)
4354 return list_entry(p
, struct kmem_cache
, next
);
4357 static inline int add_caller(unsigned long *n
, unsigned long v
)
4367 unsigned long *q
= p
+ 2 * i
;
4381 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4387 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4393 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4394 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4396 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4401 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4403 #ifdef CONFIG_KALLSYMS
4404 unsigned long offset
, size
;
4405 char modname
[MODULE_NAME_LEN
+ 1], name
[KSYM_NAME_LEN
+ 1];
4407 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4408 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4410 seq_printf(m
, " [%s]", modname
);
4414 seq_printf(m
, "%p", (void *)address
);
4417 static int leaks_show(struct seq_file
*m
, void *p
)
4419 struct kmem_cache
*cachep
= p
;
4421 struct kmem_list3
*l3
;
4423 unsigned long *n
= m
->private;
4427 if (!(cachep
->flags
& SLAB_STORE_USER
))
4429 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4432 /* OK, we can do it */
4436 for_each_online_node(node
) {
4437 l3
= cachep
->nodelists
[node
];
4442 spin_lock_irq(&l3
->list_lock
);
4444 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4445 handle_slab(n
, cachep
, slabp
);
4446 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4447 handle_slab(n
, cachep
, slabp
);
4448 spin_unlock_irq(&l3
->list_lock
);
4450 name
= cachep
->name
;
4452 /* Increase the buffer size */
4453 mutex_unlock(&cache_chain_mutex
);
4454 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4456 /* Too bad, we are really out */
4458 mutex_lock(&cache_chain_mutex
);
4461 *(unsigned long *)m
->private = n
[0] * 2;
4463 mutex_lock(&cache_chain_mutex
);
4464 /* Now make sure this entry will be retried */
4468 for (i
= 0; i
< n
[1]; i
++) {
4469 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4470 show_symbol(m
, n
[2*i
+2]);
4477 const struct seq_operations slabstats_op
= {
4478 .start
= leaks_start
,
4487 * ksize - get the actual amount of memory allocated for a given object
4488 * @objp: Pointer to the object
4490 * kmalloc may internally round up allocations and return more memory
4491 * than requested. ksize() can be used to determine the actual amount of
4492 * memory allocated. The caller may use this additional memory, even though
4493 * a smaller amount of memory was initially specified with the kmalloc call.
4494 * The caller must guarantee that objp points to a valid object previously
4495 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4496 * must not be freed during the duration of the call.
4498 size_t ksize(const void *objp
)
4500 if (unlikely(objp
== NULL
))
4503 return obj_size(virt_to_cache(objp
));