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/config.h>
90 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
111 #include <asm/uaccess.h>
112 #include <asm/cacheflush.h>
113 #include <asm/tlbflush.h>
114 #include <asm/page.h>
117 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
118 * SLAB_RED_ZONE & SLAB_POISON.
119 * 0 for faster, smaller code (especially in the critical paths).
121 * STATS - 1 to collect stats for /proc/slabinfo.
122 * 0 for faster, smaller code (especially in the critical paths).
124 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
127 #ifdef CONFIG_DEBUG_SLAB
130 #define FORCED_DEBUG 1
134 #define FORCED_DEBUG 0
137 /* Shouldn't this be in a header file somewhere? */
138 #define BYTES_PER_WORD sizeof(void *)
140 #ifndef cache_line_size
141 #define cache_line_size() L1_CACHE_BYTES
144 #ifndef ARCH_KMALLOC_MINALIGN
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
150 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
151 * Note that this flag disables some debug features.
153 #define ARCH_KMALLOC_MINALIGN 0
156 #ifndef ARCH_SLAB_MINALIGN
158 * Enforce a minimum alignment for all caches.
159 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
160 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
161 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
162 * some debug features.
164 #define ARCH_SLAB_MINALIGN 0
167 #ifndef ARCH_KMALLOC_FLAGS
168 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171 /* Legal flag mask for kmem_cache_create(). */
173 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
176 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
180 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
181 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
182 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
183 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
189 * Bufctl's are used for linking objs within a slab
192 * This implementation relies on "struct page" for locating the cache &
193 * slab an object belongs to.
194 * This allows the bufctl structure to be small (one int), but limits
195 * the number of objects a slab (not a cache) can contain when off-slab
196 * bufctls are used. The limit is the size of the largest general cache
197 * that does not use off-slab slabs.
198 * For 32bit archs with 4 kB pages, is this 56.
199 * This is not serious, as it is only for large objects, when it is unwise
200 * to have too many per slab.
201 * Note: This limit can be raised by introducing a general cache whose size
202 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 typedef unsigned int kmem_bufctl_t
;
206 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
207 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
208 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
209 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 struct list_head list
;
220 unsigned long colouroff
;
221 void *s_mem
; /* including colour offset */
222 unsigned int inuse
; /* num of objs active in slab */
224 unsigned short nodeid
;
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
241 * We assume struct slab_rcu can overlay struct slab when destroying.
244 struct rcu_head head
;
245 struct kmem_cache
*cachep
;
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
257 * The limit is stored in the per-cpu structure to reduce the data cache
264 unsigned int batchcount
;
265 unsigned int touched
;
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
271 * [0] is for gcc 2.95. It should really be [].
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
279 #define BOOT_CPUCACHE_ENTRIES 1
280 struct arraycache_init
{
281 struct array_cache cache
;
282 void *entries
[BOOT_CPUCACHE_ENTRIES
];
286 * The slab lists for all objects.
289 struct list_head slabs_partial
; /* partial list first, better asm code */
290 struct list_head slabs_full
;
291 struct list_head slabs_free
;
292 unsigned long free_objects
;
293 unsigned int free_limit
;
294 unsigned int colour_next
; /* Per-node cache coloring */
295 spinlock_t list_lock
;
296 struct array_cache
*shared
; /* shared per node */
297 struct array_cache
**alien
; /* on other nodes */
298 unsigned long next_reap
; /* updated without locking */
299 int free_touched
; /* updated without locking */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
312 * This function must be completely optimized away if a constant is passed to
313 * it. Mostly the same as what is in linux/slab.h except it returns an index.
315 static __always_inline
int index_of(const size_t size
)
317 extern void __bad_size(void);
319 if (__builtin_constant_p(size
)) {
327 #include "linux/kmalloc_sizes.h"
335 static int slab_early_init
= 1;
337 #define INDEX_AC index_of(sizeof(struct arraycache_init))
338 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
340 static void kmem_list3_init(struct kmem_list3
*parent
)
342 INIT_LIST_HEAD(&parent
->slabs_full
);
343 INIT_LIST_HEAD(&parent
->slabs_partial
);
344 INIT_LIST_HEAD(&parent
->slabs_free
);
345 parent
->shared
= NULL
;
346 parent
->alien
= NULL
;
347 parent
->colour_next
= 0;
348 spin_lock_init(&parent
->list_lock
);
349 parent
->free_objects
= 0;
350 parent
->free_touched
= 0;
353 #define MAKE_LIST(cachep, listp, slab, nodeid) \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
359 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 /* 1) per-cpu data, touched during every alloc/free */
374 struct array_cache
*array
[NR_CPUS
];
375 /* 2) Cache tunables. Protected by cache_chain_mutex */
376 unsigned int batchcount
;
380 unsigned int buffer_size
;
381 /* 3) touched by every alloc & free from the backend */
382 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
384 unsigned int flags
; /* constant flags */
385 unsigned int num
; /* # of objs per slab */
387 /* 4) cache_grow/shrink */
388 /* order of pgs per slab (2^n) */
389 unsigned int gfporder
;
391 /* force GFP flags, e.g. GFP_DMA */
394 size_t colour
; /* cache colouring range */
395 unsigned int colour_off
; /* colour offset */
396 struct kmem_cache
*slabp_cache
;
397 unsigned int slab_size
;
398 unsigned int dflags
; /* dynamic flags */
400 /* constructor func */
401 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
403 /* de-constructor func */
404 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
406 /* 5) cache creation/removal */
408 struct list_head next
;
412 unsigned long num_active
;
413 unsigned long num_allocations
;
414 unsigned long high_mark
;
416 unsigned long reaped
;
417 unsigned long errors
;
418 unsigned long max_freeable
;
419 unsigned long node_allocs
;
420 unsigned long node_frees
;
421 unsigned long node_overflow
;
429 * If debugging is enabled, then the allocator can add additional
430 * fields and/or padding to every object. buffer_size contains the total
431 * object size including these internal fields, the following two
432 * variables contain the offset to the user object and its size.
439 #define CFLGS_OFF_SLAB (0x80000000UL)
440 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
442 #define BATCHREFILL_LIMIT 16
444 * Optimization question: fewer reaps means less probability for unnessary
445 * cpucache drain/refill cycles.
447 * OTOH the cpuarrays can contain lots of objects,
448 * which could lock up otherwise freeable slabs.
450 #define REAPTIMEOUT_CPUC (2*HZ)
451 #define REAPTIMEOUT_LIST3 (4*HZ)
454 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
455 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
456 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
457 #define STATS_INC_GROWN(x) ((x)->grown++)
458 #define STATS_INC_REAPED(x) ((x)->reaped++)
459 #define STATS_SET_HIGH(x) \
461 if ((x)->num_active > (x)->high_mark) \
462 (x)->high_mark = (x)->num_active; \
464 #define STATS_INC_ERR(x) ((x)->errors++)
465 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
466 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
467 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
468 #define STATS_SET_FREEABLE(x, i) \
470 if ((x)->max_freeable < i) \
471 (x)->max_freeable = i; \
473 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
474 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
475 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
476 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
478 #define STATS_INC_ACTIVE(x) do { } while (0)
479 #define STATS_DEC_ACTIVE(x) do { } while (0)
480 #define STATS_INC_ALLOCED(x) do { } while (0)
481 #define STATS_INC_GROWN(x) do { } while (0)
482 #define STATS_INC_REAPED(x) do { } while (0)
483 #define STATS_SET_HIGH(x) do { } while (0)
484 #define STATS_INC_ERR(x) do { } while (0)
485 #define STATS_INC_NODEALLOCS(x) do { } while (0)
486 #define STATS_INC_NODEFREES(x) do { } while (0)
487 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
488 #define STATS_SET_FREEABLE(x, i) do { } while (0)
489 #define STATS_INC_ALLOCHIT(x) do { } while (0)
490 #define STATS_INC_ALLOCMISS(x) do { } while (0)
491 #define STATS_INC_FREEHIT(x) do { } while (0)
492 #define STATS_INC_FREEMISS(x) do { } while (0)
498 * memory layout of objects:
500 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
501 * the end of an object is aligned with the end of the real
502 * allocation. Catches writes behind the end of the allocation.
503 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
505 * cachep->obj_offset: The real object.
506 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
507 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
508 * [BYTES_PER_WORD long]
510 static int obj_offset(struct kmem_cache
*cachep
)
512 return cachep
->obj_offset
;
515 static int obj_size(struct kmem_cache
*cachep
)
517 return cachep
->obj_size
;
520 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
522 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
523 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
526 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
528 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
529 if (cachep
->flags
& SLAB_STORE_USER
)
530 return (unsigned long *)(objp
+ cachep
->buffer_size
-
532 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
535 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
537 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
538 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
543 #define obj_offset(x) 0
544 #define obj_size(cachep) (cachep->buffer_size)
545 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
546 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
547 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
552 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
555 #if defined(CONFIG_LARGE_ALLOCS)
556 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
557 #define MAX_GFP_ORDER 13 /* up to 32Mb */
558 #elif defined(CONFIG_MMU)
559 #define MAX_OBJ_ORDER 5 /* 32 pages */
560 #define MAX_GFP_ORDER 5 /* 32 pages */
562 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
563 #define MAX_GFP_ORDER 8 /* up to 1Mb */
567 * Do not go above this order unless 0 objects fit into the slab.
569 #define BREAK_GFP_ORDER_HI 1
570 #define BREAK_GFP_ORDER_LO 0
571 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
574 * Functions for storing/retrieving the cachep and or slab from the page
575 * allocator. These are used to find the slab an obj belongs to. With kfree(),
576 * these are used to find the cache which an obj belongs to.
578 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
580 page
->lru
.next
= (struct list_head
*)cache
;
583 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
585 if (unlikely(PageCompound(page
)))
586 page
= (struct page
*)page_private(page
);
587 BUG_ON(!PageSlab(page
));
588 return (struct kmem_cache
*)page
->lru
.next
;
591 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
593 page
->lru
.prev
= (struct list_head
*)slab
;
596 static inline struct slab
*page_get_slab(struct page
*page
)
598 if (unlikely(PageCompound(page
)))
599 page
= (struct page
*)page_private(page
);
600 BUG_ON(!PageSlab(page
));
601 return (struct slab
*)page
->lru
.prev
;
604 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
606 struct page
*page
= virt_to_page(obj
);
607 return page_get_cache(page
);
610 static inline struct slab
*virt_to_slab(const void *obj
)
612 struct page
*page
= virt_to_page(obj
);
613 return page_get_slab(page
);
616 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
619 return slab
->s_mem
+ cache
->buffer_size
* idx
;
622 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
623 struct slab
*slab
, void *obj
)
625 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
629 * These are the default caches for kmalloc. Custom caches can have other sizes.
631 struct cache_sizes malloc_sizes
[] = {
632 #define CACHE(x) { .cs_size = (x) },
633 #include <linux/kmalloc_sizes.h>
637 EXPORT_SYMBOL(malloc_sizes
);
639 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
645 static struct cache_names __initdata cache_names
[] = {
646 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
647 #include <linux/kmalloc_sizes.h>
652 static struct arraycache_init initarray_cache __initdata
=
653 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
654 static struct arraycache_init initarray_generic
=
655 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
657 /* internal cache of cache description objs */
658 static struct kmem_cache cache_cache
= {
660 .limit
= BOOT_CPUCACHE_ENTRIES
,
662 .buffer_size
= sizeof(struct kmem_cache
),
663 .name
= "kmem_cache",
665 .obj_size
= sizeof(struct kmem_cache
),
669 /* Guard access to the cache-chain. */
670 static DEFINE_MUTEX(cache_chain_mutex
);
671 static struct list_head cache_chain
;
674 * vm_enough_memory() looks at this to determine how many slab-allocated pages
675 * are possibly freeable under pressure
677 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
679 atomic_t slab_reclaim_pages
;
682 * chicken and egg problem: delay the per-cpu array allocation
683 * until the general caches are up.
693 * used by boot code to determine if it can use slab based allocator
695 int slab_is_available(void)
697 return g_cpucache_up
== FULL
;
700 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
702 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
704 static void enable_cpucache(struct kmem_cache
*cachep
);
705 static void cache_reap(void *unused
);
706 static int __node_shrink(struct kmem_cache
*cachep
, int node
);
708 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
710 return cachep
->array
[smp_processor_id()];
713 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
716 struct cache_sizes
*csizep
= malloc_sizes
;
719 /* This happens if someone tries to call
720 * kmem_cache_create(), or __kmalloc(), before
721 * the generic caches are initialized.
723 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
725 while (size
> csizep
->cs_size
)
729 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
730 * has cs_{dma,}cachep==NULL. Thus no special case
731 * for large kmalloc calls required.
733 if (unlikely(gfpflags
& GFP_DMA
))
734 return csizep
->cs_dmacachep
;
735 return csizep
->cs_cachep
;
738 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
740 return __find_general_cachep(size
, gfpflags
);
742 EXPORT_SYMBOL(kmem_find_general_cachep
);
744 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
746 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
750 * Calculate the number of objects and left-over bytes for a given buffer size.
752 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
753 size_t align
, int flags
, size_t *left_over
,
758 size_t slab_size
= PAGE_SIZE
<< gfporder
;
761 * The slab management structure can be either off the slab or
762 * on it. For the latter case, the memory allocated for a
766 * - One kmem_bufctl_t for each object
767 * - Padding to respect alignment of @align
768 * - @buffer_size bytes for each object
770 * If the slab management structure is off the slab, then the
771 * alignment will already be calculated into the size. Because
772 * the slabs are all pages aligned, the objects will be at the
773 * correct alignment when allocated.
775 if (flags
& CFLGS_OFF_SLAB
) {
777 nr_objs
= slab_size
/ buffer_size
;
779 if (nr_objs
> SLAB_LIMIT
)
780 nr_objs
= SLAB_LIMIT
;
783 * Ignore padding for the initial guess. The padding
784 * is at most @align-1 bytes, and @buffer_size is at
785 * least @align. In the worst case, this result will
786 * be one greater than the number of objects that fit
787 * into the memory allocation when taking the padding
790 nr_objs
= (slab_size
- sizeof(struct slab
)) /
791 (buffer_size
+ sizeof(kmem_bufctl_t
));
794 * This calculated number will be either the right
795 * amount, or one greater than what we want.
797 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
801 if (nr_objs
> SLAB_LIMIT
)
802 nr_objs
= SLAB_LIMIT
;
804 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
807 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
810 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
812 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
815 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
816 function
, cachep
->name
, msg
);
822 * Special reaping functions for NUMA systems called from cache_reap().
823 * These take care of doing round robin flushing of alien caches (containing
824 * objects freed on different nodes from which they were allocated) and the
825 * flushing of remote pcps by calling drain_node_pages.
827 static DEFINE_PER_CPU(unsigned long, reap_node
);
829 static void init_reap_node(int cpu
)
833 node
= next_node(cpu_to_node(cpu
), node_online_map
);
834 if (node
== MAX_NUMNODES
)
835 node
= first_node(node_online_map
);
837 __get_cpu_var(reap_node
) = node
;
840 static void next_reap_node(void)
842 int node
= __get_cpu_var(reap_node
);
845 * Also drain per cpu pages on remote zones
847 if (node
!= numa_node_id())
848 drain_node_pages(node
);
850 node
= next_node(node
, node_online_map
);
851 if (unlikely(node
>= MAX_NUMNODES
))
852 node
= first_node(node_online_map
);
853 __get_cpu_var(reap_node
) = node
;
857 #define init_reap_node(cpu) do { } while (0)
858 #define next_reap_node(void) do { } while (0)
862 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
863 * via the workqueue/eventd.
864 * Add the CPU number into the expiration time to minimize the possibility of
865 * the CPUs getting into lockstep and contending for the global cache chain
868 static void __devinit
start_cpu_timer(int cpu
)
870 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
873 * When this gets called from do_initcalls via cpucache_init(),
874 * init_workqueues() has already run, so keventd will be setup
877 if (keventd_up() && reap_work
->func
== NULL
) {
879 INIT_WORK(reap_work
, cache_reap
, NULL
);
880 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
884 static struct array_cache
*alloc_arraycache(int node
, int entries
,
887 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
888 struct array_cache
*nc
= NULL
;
890 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
894 nc
->batchcount
= batchcount
;
896 spin_lock_init(&nc
->lock
);
902 * Transfer objects in one arraycache to another.
903 * Locking must be handled by the caller.
905 * Return the number of entries transferred.
907 static int transfer_objects(struct array_cache
*to
,
908 struct array_cache
*from
, unsigned int max
)
910 /* Figure out how many entries to transfer */
911 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
916 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
926 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
927 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
929 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
931 struct array_cache
**ac_ptr
;
932 int memsize
= sizeof(void *) * MAX_NUMNODES
;
937 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
940 if (i
== node
|| !node_online(i
)) {
944 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
946 for (i
--; i
<= 0; i
--)
956 static void free_alien_cache(struct array_cache
**ac_ptr
)
967 static void __drain_alien_cache(struct kmem_cache
*cachep
,
968 struct array_cache
*ac
, int node
)
970 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
973 spin_lock(&rl3
->list_lock
);
975 * Stuff objects into the remote nodes shared array first.
976 * That way we could avoid the overhead of putting the objects
977 * into the free lists and getting them back later.
980 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
982 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
984 spin_unlock(&rl3
->list_lock
);
989 * Called from cache_reap() to regularly drain alien caches round robin.
991 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
993 int node
= __get_cpu_var(reap_node
);
996 struct array_cache
*ac
= l3
->alien
[node
];
998 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
999 __drain_alien_cache(cachep
, ac
, node
);
1000 spin_unlock_irq(&ac
->lock
);
1005 static void drain_alien_cache(struct kmem_cache
*cachep
,
1006 struct array_cache
**alien
)
1009 struct array_cache
*ac
;
1010 unsigned long flags
;
1012 for_each_online_node(i
) {
1015 spin_lock_irqsave(&ac
->lock
, flags
);
1016 __drain_alien_cache(cachep
, ac
, i
);
1017 spin_unlock_irqrestore(&ac
->lock
, flags
);
1022 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1024 struct slab
*slabp
= virt_to_slab(objp
);
1025 int nodeid
= slabp
->nodeid
;
1026 struct kmem_list3
*l3
;
1027 struct array_cache
*alien
= NULL
;
1030 * Make sure we are not freeing a object from another node to the array
1031 * cache on this cpu.
1033 if (likely(slabp
->nodeid
== numa_node_id()))
1036 l3
= cachep
->nodelists
[numa_node_id()];
1037 STATS_INC_NODEFREES(cachep
);
1038 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1039 alien
= l3
->alien
[nodeid
];
1040 spin_lock(&alien
->lock
);
1041 if (unlikely(alien
->avail
== alien
->limit
)) {
1042 STATS_INC_ACOVERFLOW(cachep
);
1043 __drain_alien_cache(cachep
, alien
, nodeid
);
1045 alien
->entry
[alien
->avail
++] = objp
;
1046 spin_unlock(&alien
->lock
);
1048 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1049 free_block(cachep
, &objp
, 1, nodeid
);
1050 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1057 #define drain_alien_cache(cachep, alien) do { } while (0)
1058 #define reap_alien(cachep, l3) do { } while (0)
1060 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1062 return (struct array_cache
**) 0x01020304ul
;
1065 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1069 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1076 static int __devinit
cpuup_callback(struct notifier_block
*nfb
,
1077 unsigned long action
, void *hcpu
)
1079 long cpu
= (long)hcpu
;
1080 struct kmem_cache
*cachep
;
1081 struct kmem_list3
*l3
= NULL
;
1082 int node
= cpu_to_node(cpu
);
1083 int memsize
= sizeof(struct kmem_list3
);
1086 case CPU_UP_PREPARE
:
1087 mutex_lock(&cache_chain_mutex
);
1089 * We need to do this right in the beginning since
1090 * alloc_arraycache's are going to use this list.
1091 * kmalloc_node allows us to add the slab to the right
1092 * kmem_list3 and not this cpu's kmem_list3
1095 list_for_each_entry(cachep
, &cache_chain
, next
) {
1097 * Set up the size64 kmemlist for cpu before we can
1098 * begin anything. Make sure some other cpu on this
1099 * node has not already allocated this
1101 if (!cachep
->nodelists
[node
]) {
1102 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1105 kmem_list3_init(l3
);
1106 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1107 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1110 * The l3s don't come and go as CPUs come and
1111 * go. cache_chain_mutex is sufficient
1114 cachep
->nodelists
[node
] = l3
;
1117 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1118 cachep
->nodelists
[node
]->free_limit
=
1119 (1 + nr_cpus_node(node
)) *
1120 cachep
->batchcount
+ cachep
->num
;
1121 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1125 * Now we can go ahead with allocating the shared arrays and
1128 list_for_each_entry(cachep
, &cache_chain
, next
) {
1129 struct array_cache
*nc
;
1130 struct array_cache
*shared
;
1131 struct array_cache
**alien
;
1133 nc
= alloc_arraycache(node
, cachep
->limit
,
1134 cachep
->batchcount
);
1137 shared
= alloc_arraycache(node
,
1138 cachep
->shared
* cachep
->batchcount
,
1143 alien
= alloc_alien_cache(node
, cachep
->limit
);
1146 cachep
->array
[cpu
] = nc
;
1147 l3
= cachep
->nodelists
[node
];
1150 spin_lock_irq(&l3
->list_lock
);
1153 * We are serialised from CPU_DEAD or
1154 * CPU_UP_CANCELLED by the cpucontrol lock
1156 l3
->shared
= shared
;
1165 spin_unlock_irq(&l3
->list_lock
);
1167 free_alien_cache(alien
);
1169 mutex_unlock(&cache_chain_mutex
);
1172 start_cpu_timer(cpu
);
1174 #ifdef CONFIG_HOTPLUG_CPU
1177 * Even if all the cpus of a node are down, we don't free the
1178 * kmem_list3 of any cache. This to avoid a race between
1179 * cpu_down, and a kmalloc allocation from another cpu for
1180 * memory from the node of the cpu going down. The list3
1181 * structure is usually allocated from kmem_cache_create() and
1182 * gets destroyed at kmem_cache_destroy().
1185 case CPU_UP_CANCELED
:
1186 mutex_lock(&cache_chain_mutex
);
1187 list_for_each_entry(cachep
, &cache_chain
, next
) {
1188 struct array_cache
*nc
;
1189 struct array_cache
*shared
;
1190 struct array_cache
**alien
;
1193 mask
= node_to_cpumask(node
);
1194 /* cpu is dead; no one can alloc from it. */
1195 nc
= cachep
->array
[cpu
];
1196 cachep
->array
[cpu
] = NULL
;
1197 l3
= cachep
->nodelists
[node
];
1200 goto free_array_cache
;
1202 spin_lock_irq(&l3
->list_lock
);
1204 /* Free limit for this kmem_list3 */
1205 l3
->free_limit
-= cachep
->batchcount
;
1207 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1209 if (!cpus_empty(mask
)) {
1210 spin_unlock_irq(&l3
->list_lock
);
1211 goto free_array_cache
;
1214 shared
= l3
->shared
;
1216 free_block(cachep
, l3
->shared
->entry
,
1217 l3
->shared
->avail
, node
);
1224 spin_unlock_irq(&l3
->list_lock
);
1228 drain_alien_cache(cachep
, alien
);
1229 free_alien_cache(alien
);
1235 * In the previous loop, all the objects were freed to
1236 * the respective cache's slabs, now we can go ahead and
1237 * shrink each nodelist to its limit.
1239 list_for_each_entry(cachep
, &cache_chain
, next
) {
1240 l3
= cachep
->nodelists
[node
];
1243 spin_lock_irq(&l3
->list_lock
);
1244 /* free slabs belonging to this node */
1245 __node_shrink(cachep
, node
);
1246 spin_unlock_irq(&l3
->list_lock
);
1248 mutex_unlock(&cache_chain_mutex
);
1254 mutex_unlock(&cache_chain_mutex
);
1258 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
1261 * swap the static kmem_list3 with kmalloced memory
1263 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1266 struct kmem_list3
*ptr
;
1268 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1269 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1272 local_irq_disable();
1273 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1274 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1275 cachep
->nodelists
[nodeid
] = ptr
;
1280 * Initialisation. Called after the page allocator have been initialised and
1281 * before smp_init().
1283 void __init
kmem_cache_init(void)
1286 struct cache_sizes
*sizes
;
1287 struct cache_names
*names
;
1291 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1292 kmem_list3_init(&initkmem_list3
[i
]);
1293 if (i
< MAX_NUMNODES
)
1294 cache_cache
.nodelists
[i
] = NULL
;
1298 * Fragmentation resistance on low memory - only use bigger
1299 * page orders on machines with more than 32MB of memory.
1301 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1302 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1304 /* Bootstrap is tricky, because several objects are allocated
1305 * from caches that do not exist yet:
1306 * 1) initialize the cache_cache cache: it contains the struct
1307 * kmem_cache structures of all caches, except cache_cache itself:
1308 * cache_cache is statically allocated.
1309 * Initially an __init data area is used for the head array and the
1310 * kmem_list3 structures, it's replaced with a kmalloc allocated
1311 * array at the end of the bootstrap.
1312 * 2) Create the first kmalloc cache.
1313 * The struct kmem_cache for the new cache is allocated normally.
1314 * An __init data area is used for the head array.
1315 * 3) Create the remaining kmalloc caches, with minimally sized
1317 * 4) Replace the __init data head arrays for cache_cache and the first
1318 * kmalloc cache with kmalloc allocated arrays.
1319 * 5) Replace the __init data for kmem_list3 for cache_cache and
1320 * the other cache's with kmalloc allocated memory.
1321 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1324 /* 1) create the cache_cache */
1325 INIT_LIST_HEAD(&cache_chain
);
1326 list_add(&cache_cache
.next
, &cache_chain
);
1327 cache_cache
.colour_off
= cache_line_size();
1328 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1329 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1331 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1334 for (order
= 0; order
< MAX_ORDER
; order
++) {
1335 cache_estimate(order
, cache_cache
.buffer_size
,
1336 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1337 if (cache_cache
.num
)
1340 BUG_ON(!cache_cache
.num
);
1341 cache_cache
.gfporder
= order
;
1342 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1343 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1344 sizeof(struct slab
), cache_line_size());
1346 /* 2+3) create the kmalloc caches */
1347 sizes
= malloc_sizes
;
1348 names
= cache_names
;
1351 * Initialize the caches that provide memory for the array cache and the
1352 * kmem_list3 structures first. Without this, further allocations will
1356 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1357 sizes
[INDEX_AC
].cs_size
,
1358 ARCH_KMALLOC_MINALIGN
,
1359 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1362 if (INDEX_AC
!= INDEX_L3
) {
1363 sizes
[INDEX_L3
].cs_cachep
=
1364 kmem_cache_create(names
[INDEX_L3
].name
,
1365 sizes
[INDEX_L3
].cs_size
,
1366 ARCH_KMALLOC_MINALIGN
,
1367 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1371 slab_early_init
= 0;
1373 while (sizes
->cs_size
!= ULONG_MAX
) {
1375 * For performance, all the general caches are L1 aligned.
1376 * This should be particularly beneficial on SMP boxes, as it
1377 * eliminates "false sharing".
1378 * Note for systems short on memory removing the alignment will
1379 * allow tighter packing of the smaller caches.
1381 if (!sizes
->cs_cachep
) {
1382 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1384 ARCH_KMALLOC_MINALIGN
,
1385 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1389 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1391 ARCH_KMALLOC_MINALIGN
,
1392 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1398 /* 4) Replace the bootstrap head arrays */
1402 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1404 local_irq_disable();
1405 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1406 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1407 sizeof(struct arraycache_init
));
1408 cache_cache
.array
[smp_processor_id()] = ptr
;
1411 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1413 local_irq_disable();
1414 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1415 != &initarray_generic
.cache
);
1416 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1417 sizeof(struct arraycache_init
));
1418 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1422 /* 5) Replace the bootstrap kmem_list3's */
1425 /* Replace the static kmem_list3 structures for the boot cpu */
1426 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1429 for_each_online_node(node
) {
1430 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1431 &initkmem_list3
[SIZE_AC
+ node
], node
);
1433 if (INDEX_AC
!= INDEX_L3
) {
1434 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1435 &initkmem_list3
[SIZE_L3
+ node
],
1441 /* 6) resize the head arrays to their final sizes */
1443 struct kmem_cache
*cachep
;
1444 mutex_lock(&cache_chain_mutex
);
1445 list_for_each_entry(cachep
, &cache_chain
, next
)
1446 enable_cpucache(cachep
);
1447 mutex_unlock(&cache_chain_mutex
);
1451 g_cpucache_up
= FULL
;
1454 * Register a cpu startup notifier callback that initializes
1455 * cpu_cache_get for all new cpus
1457 register_cpu_notifier(&cpucache_notifier
);
1460 * The reap timers are started later, with a module init call: That part
1461 * of the kernel is not yet operational.
1465 static int __init
cpucache_init(void)
1470 * Register the timers that return unneeded pages to the page allocator
1472 for_each_online_cpu(cpu
)
1473 start_cpu_timer(cpu
);
1476 __initcall(cpucache_init
);
1479 * Interface to system's page allocator. No need to hold the cache-lock.
1481 * If we requested dmaable memory, we will get it. Even if we
1482 * did not request dmaable memory, we might get it, but that
1483 * would be relatively rare and ignorable.
1485 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1493 * Nommu uses slab's for process anonymous memory allocations, and thus
1494 * requires __GFP_COMP to properly refcount higher order allocations
1496 flags
|= __GFP_COMP
;
1498 flags
|= cachep
->gfpflags
;
1500 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1504 nr_pages
= (1 << cachep
->gfporder
);
1505 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1506 atomic_add(nr_pages
, &slab_reclaim_pages
);
1507 add_page_state(nr_slab
, nr_pages
);
1508 for (i
= 0; i
< nr_pages
; i
++)
1509 __SetPageSlab(page
+ i
);
1510 return page_address(page
);
1514 * Interface to system's page release.
1516 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1518 unsigned long i
= (1 << cachep
->gfporder
);
1519 struct page
*page
= virt_to_page(addr
);
1520 const unsigned long nr_freed
= i
;
1523 BUG_ON(!PageSlab(page
));
1524 __ClearPageSlab(page
);
1527 sub_page_state(nr_slab
, nr_freed
);
1528 if (current
->reclaim_state
)
1529 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1530 free_pages((unsigned long)addr
, cachep
->gfporder
);
1531 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1532 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1535 static void kmem_rcu_free(struct rcu_head
*head
)
1537 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1538 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1540 kmem_freepages(cachep
, slab_rcu
->addr
);
1541 if (OFF_SLAB(cachep
))
1542 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1547 #ifdef CONFIG_DEBUG_PAGEALLOC
1548 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1549 unsigned long caller
)
1551 int size
= obj_size(cachep
);
1553 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1555 if (size
< 5 * sizeof(unsigned long))
1558 *addr
++ = 0x12345678;
1560 *addr
++ = smp_processor_id();
1561 size
-= 3 * sizeof(unsigned long);
1563 unsigned long *sptr
= &caller
;
1564 unsigned long svalue
;
1566 while (!kstack_end(sptr
)) {
1568 if (kernel_text_address(svalue
)) {
1570 size
-= sizeof(unsigned long);
1571 if (size
<= sizeof(unsigned long))
1577 *addr
++ = 0x87654321;
1581 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1583 int size
= obj_size(cachep
);
1584 addr
= &((char *)addr
)[obj_offset(cachep
)];
1586 memset(addr
, val
, size
);
1587 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1590 static void dump_line(char *data
, int offset
, int limit
)
1593 printk(KERN_ERR
"%03x:", offset
);
1594 for (i
= 0; i
< limit
; i
++)
1595 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1602 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1607 if (cachep
->flags
& SLAB_RED_ZONE
) {
1608 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1609 *dbg_redzone1(cachep
, objp
),
1610 *dbg_redzone2(cachep
, objp
));
1613 if (cachep
->flags
& SLAB_STORE_USER
) {
1614 printk(KERN_ERR
"Last user: [<%p>]",
1615 *dbg_userword(cachep
, objp
));
1616 print_symbol("(%s)",
1617 (unsigned long)*dbg_userword(cachep
, objp
));
1620 realobj
= (char *)objp
+ obj_offset(cachep
);
1621 size
= obj_size(cachep
);
1622 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1625 if (i
+ limit
> size
)
1627 dump_line(realobj
, i
, limit
);
1631 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1637 realobj
= (char *)objp
+ obj_offset(cachep
);
1638 size
= obj_size(cachep
);
1640 for (i
= 0; i
< size
; i
++) {
1641 char exp
= POISON_FREE
;
1644 if (realobj
[i
] != exp
) {
1650 "Slab corruption: start=%p, len=%d\n",
1652 print_objinfo(cachep
, objp
, 0);
1654 /* Hexdump the affected line */
1657 if (i
+ limit
> size
)
1659 dump_line(realobj
, i
, limit
);
1662 /* Limit to 5 lines */
1668 /* Print some data about the neighboring objects, if they
1671 struct slab
*slabp
= virt_to_slab(objp
);
1674 objnr
= obj_to_index(cachep
, slabp
, objp
);
1676 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1677 realobj
= (char *)objp
+ obj_offset(cachep
);
1678 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1680 print_objinfo(cachep
, objp
, 2);
1682 if (objnr
+ 1 < cachep
->num
) {
1683 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1684 realobj
= (char *)objp
+ obj_offset(cachep
);
1685 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1687 print_objinfo(cachep
, objp
, 2);
1695 * slab_destroy_objs - destroy a slab and its objects
1696 * @cachep: cache pointer being destroyed
1697 * @slabp: slab pointer being destroyed
1699 * Call the registered destructor for each object in a slab that is being
1702 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1705 for (i
= 0; i
< cachep
->num
; i
++) {
1706 void *objp
= index_to_obj(cachep
, slabp
, i
);
1708 if (cachep
->flags
& SLAB_POISON
) {
1709 #ifdef CONFIG_DEBUG_PAGEALLOC
1710 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1712 kernel_map_pages(virt_to_page(objp
),
1713 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1715 check_poison_obj(cachep
, objp
);
1717 check_poison_obj(cachep
, objp
);
1720 if (cachep
->flags
& SLAB_RED_ZONE
) {
1721 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1722 slab_error(cachep
, "start of a freed object "
1724 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1725 slab_error(cachep
, "end of a freed object "
1728 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1729 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1733 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1737 for (i
= 0; i
< cachep
->num
; i
++) {
1738 void *objp
= index_to_obj(cachep
, slabp
, i
);
1739 (cachep
->dtor
) (objp
, cachep
, 0);
1746 * slab_destroy - destroy and release all objects in a slab
1747 * @cachep: cache pointer being destroyed
1748 * @slabp: slab pointer being destroyed
1750 * Destroy all the objs in a slab, and release the mem back to the system.
1751 * Before calling the slab must have been unlinked from the cache. The
1752 * cache-lock is not held/needed.
1754 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1756 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1758 slab_destroy_objs(cachep
, slabp
);
1759 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1760 struct slab_rcu
*slab_rcu
;
1762 slab_rcu
= (struct slab_rcu
*)slabp
;
1763 slab_rcu
->cachep
= cachep
;
1764 slab_rcu
->addr
= addr
;
1765 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1767 kmem_freepages(cachep
, addr
);
1768 if (OFF_SLAB(cachep
))
1769 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1774 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1775 * size of kmem_list3.
1777 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1781 for_each_online_node(node
) {
1782 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1783 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1785 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1790 * calculate_slab_order - calculate size (page order) of slabs
1791 * @cachep: pointer to the cache that is being created
1792 * @size: size of objects to be created in this cache.
1793 * @align: required alignment for the objects.
1794 * @flags: slab allocation flags
1796 * Also calculates the number of objects per slab.
1798 * This could be made much more intelligent. For now, try to avoid using
1799 * high order pages for slabs. When the gfp() functions are more friendly
1800 * towards high-order requests, this should be changed.
1802 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1803 size_t size
, size_t align
, unsigned long flags
)
1805 unsigned long offslab_limit
;
1806 size_t left_over
= 0;
1809 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1813 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1817 if (flags
& CFLGS_OFF_SLAB
) {
1819 * Max number of objs-per-slab for caches which
1820 * use off-slab slabs. Needed to avoid a possible
1821 * looping condition in cache_grow().
1823 offslab_limit
= size
- sizeof(struct slab
);
1824 offslab_limit
/= sizeof(kmem_bufctl_t
);
1826 if (num
> offslab_limit
)
1830 /* Found something acceptable - save it away */
1832 cachep
->gfporder
= gfporder
;
1833 left_over
= remainder
;
1836 * A VFS-reclaimable slab tends to have most allocations
1837 * as GFP_NOFS and we really don't want to have to be allocating
1838 * higher-order pages when we are unable to shrink dcache.
1840 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1844 * Large number of objects is good, but very large slabs are
1845 * currently bad for the gfp()s.
1847 if (gfporder
>= slab_break_gfp_order
)
1851 * Acceptable internal fragmentation?
1853 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1859 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1861 if (g_cpucache_up
== FULL
) {
1862 enable_cpucache(cachep
);
1865 if (g_cpucache_up
== NONE
) {
1867 * Note: the first kmem_cache_create must create the cache
1868 * that's used by kmalloc(24), otherwise the creation of
1869 * further caches will BUG().
1871 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1874 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1875 * the first cache, then we need to set up all its list3s,
1876 * otherwise the creation of further caches will BUG().
1878 set_up_list3s(cachep
, SIZE_AC
);
1879 if (INDEX_AC
== INDEX_L3
)
1880 g_cpucache_up
= PARTIAL_L3
;
1882 g_cpucache_up
= PARTIAL_AC
;
1884 cachep
->array
[smp_processor_id()] =
1885 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1887 if (g_cpucache_up
== PARTIAL_AC
) {
1888 set_up_list3s(cachep
, SIZE_L3
);
1889 g_cpucache_up
= PARTIAL_L3
;
1892 for_each_online_node(node
) {
1893 cachep
->nodelists
[node
] =
1894 kmalloc_node(sizeof(struct kmem_list3
),
1896 BUG_ON(!cachep
->nodelists
[node
]);
1897 kmem_list3_init(cachep
->nodelists
[node
]);
1901 cachep
->nodelists
[numa_node_id()]->next_reap
=
1902 jiffies
+ REAPTIMEOUT_LIST3
+
1903 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1905 cpu_cache_get(cachep
)->avail
= 0;
1906 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1907 cpu_cache_get(cachep
)->batchcount
= 1;
1908 cpu_cache_get(cachep
)->touched
= 0;
1909 cachep
->batchcount
= 1;
1910 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1914 * kmem_cache_create - Create a cache.
1915 * @name: A string which is used in /proc/slabinfo to identify this cache.
1916 * @size: The size of objects to be created in this cache.
1917 * @align: The required alignment for the objects.
1918 * @flags: SLAB flags
1919 * @ctor: A constructor for the objects.
1920 * @dtor: A destructor for the objects.
1922 * Returns a ptr to the cache on success, NULL on failure.
1923 * Cannot be called within a int, but can be interrupted.
1924 * The @ctor is run when new pages are allocated by the cache
1925 * and the @dtor is run before the pages are handed back.
1927 * @name must be valid until the cache is destroyed. This implies that
1928 * the module calling this has to destroy the cache before getting unloaded.
1932 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1933 * to catch references to uninitialised memory.
1935 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1936 * for buffer overruns.
1938 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1939 * cacheline. This can be beneficial if you're counting cycles as closely
1943 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1944 unsigned long flags
,
1945 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1946 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1948 size_t left_over
, slab_size
, ralign
;
1949 struct kmem_cache
*cachep
= NULL
, *pc
;
1952 * Sanity checks... these are all serious usage bugs.
1954 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
1955 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1956 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
1962 * Prevent CPUs from coming and going.
1963 * lock_cpu_hotplug() nests outside cache_chain_mutex
1967 mutex_lock(&cache_chain_mutex
);
1969 list_for_each_entry(pc
, &cache_chain
, next
) {
1970 mm_segment_t old_fs
= get_fs();
1975 * This happens when the module gets unloaded and doesn't
1976 * destroy its slab cache and no-one else reuses the vmalloc
1977 * area of the module. Print a warning.
1980 res
= __get_user(tmp
, pc
->name
);
1983 printk("SLAB: cache with size %d has lost its name\n",
1988 if (!strcmp(pc
->name
, name
)) {
1989 printk("kmem_cache_create: duplicate cache %s\n", name
);
1996 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1997 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1998 /* No constructor, but inital state check requested */
1999 printk(KERN_ERR
"%s: No con, but init state check "
2000 "requested - %s\n", __FUNCTION__
, name
);
2001 flags
&= ~SLAB_DEBUG_INITIAL
;
2005 * Enable redzoning and last user accounting, except for caches with
2006 * large objects, if the increased size would increase the object size
2007 * above the next power of two: caches with object sizes just above a
2008 * power of two have a significant amount of internal fragmentation.
2010 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2011 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2012 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2013 flags
|= SLAB_POISON
;
2015 if (flags
& SLAB_DESTROY_BY_RCU
)
2016 BUG_ON(flags
& SLAB_POISON
);
2018 if (flags
& SLAB_DESTROY_BY_RCU
)
2022 * Always checks flags, a caller might be expecting debug support which
2025 BUG_ON(flags
& ~CREATE_MASK
);
2028 * Check that size is in terms of words. This is needed to avoid
2029 * unaligned accesses for some archs when redzoning is used, and makes
2030 * sure any on-slab bufctl's are also correctly aligned.
2032 if (size
& (BYTES_PER_WORD
- 1)) {
2033 size
+= (BYTES_PER_WORD
- 1);
2034 size
&= ~(BYTES_PER_WORD
- 1);
2037 /* calculate the final buffer alignment: */
2039 /* 1) arch recommendation: can be overridden for debug */
2040 if (flags
& SLAB_HWCACHE_ALIGN
) {
2042 * Default alignment: as specified by the arch code. Except if
2043 * an object is really small, then squeeze multiple objects into
2046 ralign
= cache_line_size();
2047 while (size
<= ralign
/ 2)
2050 ralign
= BYTES_PER_WORD
;
2052 /* 2) arch mandated alignment: disables debug if necessary */
2053 if (ralign
< ARCH_SLAB_MINALIGN
) {
2054 ralign
= ARCH_SLAB_MINALIGN
;
2055 if (ralign
> BYTES_PER_WORD
)
2056 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2058 /* 3) caller mandated alignment: disables debug if necessary */
2059 if (ralign
< align
) {
2061 if (ralign
> BYTES_PER_WORD
)
2062 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2065 * 4) Store it. Note that the debug code below can reduce
2066 * the alignment to BYTES_PER_WORD.
2070 /* Get cache's description obj. */
2071 cachep
= kmem_cache_zalloc(&cache_cache
, SLAB_KERNEL
);
2076 cachep
->obj_size
= size
;
2078 if (flags
& SLAB_RED_ZONE
) {
2079 /* redzoning only works with word aligned caches */
2080 align
= BYTES_PER_WORD
;
2082 /* add space for red zone words */
2083 cachep
->obj_offset
+= BYTES_PER_WORD
;
2084 size
+= 2 * BYTES_PER_WORD
;
2086 if (flags
& SLAB_STORE_USER
) {
2087 /* user store requires word alignment and
2088 * one word storage behind the end of the real
2091 align
= BYTES_PER_WORD
;
2092 size
+= BYTES_PER_WORD
;
2094 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2095 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2096 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2097 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2104 * Determine if the slab management is 'on' or 'off' slab.
2105 * (bootstrapping cannot cope with offslab caches so don't do
2108 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2110 * Size is large, assume best to place the slab management obj
2111 * off-slab (should allow better packing of objs).
2113 flags
|= CFLGS_OFF_SLAB
;
2115 size
= ALIGN(size
, align
);
2117 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2120 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2121 kmem_cache_free(&cache_cache
, cachep
);
2125 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2126 + sizeof(struct slab
), align
);
2129 * If the slab has been placed off-slab, and we have enough space then
2130 * move it on-slab. This is at the expense of any extra colouring.
2132 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2133 flags
&= ~CFLGS_OFF_SLAB
;
2134 left_over
-= slab_size
;
2137 if (flags
& CFLGS_OFF_SLAB
) {
2138 /* really off slab. No need for manual alignment */
2140 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2143 cachep
->colour_off
= cache_line_size();
2144 /* Offset must be a multiple of the alignment. */
2145 if (cachep
->colour_off
< align
)
2146 cachep
->colour_off
= align
;
2147 cachep
->colour
= left_over
/ cachep
->colour_off
;
2148 cachep
->slab_size
= slab_size
;
2149 cachep
->flags
= flags
;
2150 cachep
->gfpflags
= 0;
2151 if (flags
& SLAB_CACHE_DMA
)
2152 cachep
->gfpflags
|= GFP_DMA
;
2153 cachep
->buffer_size
= size
;
2155 if (flags
& CFLGS_OFF_SLAB
)
2156 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2157 cachep
->ctor
= ctor
;
2158 cachep
->dtor
= dtor
;
2159 cachep
->name
= name
;
2162 setup_cpu_cache(cachep
);
2164 /* cache setup completed, link it into the list */
2165 list_add(&cachep
->next
, &cache_chain
);
2167 if (!cachep
&& (flags
& SLAB_PANIC
))
2168 panic("kmem_cache_create(): failed to create slab `%s'\n",
2170 mutex_unlock(&cache_chain_mutex
);
2171 unlock_cpu_hotplug();
2174 EXPORT_SYMBOL(kmem_cache_create
);
2177 static void check_irq_off(void)
2179 BUG_ON(!irqs_disabled());
2182 static void check_irq_on(void)
2184 BUG_ON(irqs_disabled());
2187 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2191 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2195 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2199 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2204 #define check_irq_off() do { } while(0)
2205 #define check_irq_on() do { } while(0)
2206 #define check_spinlock_acquired(x) do { } while(0)
2207 #define check_spinlock_acquired_node(x, y) do { } while(0)
2210 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2211 struct array_cache
*ac
,
2212 int force
, int node
);
2214 static void do_drain(void *arg
)
2216 struct kmem_cache
*cachep
= arg
;
2217 struct array_cache
*ac
;
2218 int node
= numa_node_id();
2221 ac
= cpu_cache_get(cachep
);
2222 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2223 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2224 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2228 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2230 struct kmem_list3
*l3
;
2233 on_each_cpu(do_drain
, cachep
, 1, 1);
2235 for_each_online_node(node
) {
2236 l3
= cachep
->nodelists
[node
];
2237 if (l3
&& l3
->alien
)
2238 drain_alien_cache(cachep
, l3
->alien
);
2241 for_each_online_node(node
) {
2242 l3
= cachep
->nodelists
[node
];
2244 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2248 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2251 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2255 struct list_head
*p
;
2257 p
= l3
->slabs_free
.prev
;
2258 if (p
== &l3
->slabs_free
)
2261 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2263 BUG_ON(slabp
->inuse
);
2265 list_del(&slabp
->list
);
2267 l3
->free_objects
-= cachep
->num
;
2268 spin_unlock_irq(&l3
->list_lock
);
2269 slab_destroy(cachep
, slabp
);
2270 spin_lock_irq(&l3
->list_lock
);
2272 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2276 static int __cache_shrink(struct kmem_cache
*cachep
)
2279 struct kmem_list3
*l3
;
2281 drain_cpu_caches(cachep
);
2284 for_each_online_node(i
) {
2285 l3
= cachep
->nodelists
[i
];
2287 spin_lock_irq(&l3
->list_lock
);
2288 ret
+= __node_shrink(cachep
, i
);
2289 spin_unlock_irq(&l3
->list_lock
);
2292 return (ret
? 1 : 0);
2296 * kmem_cache_shrink - Shrink a cache.
2297 * @cachep: The cache to shrink.
2299 * Releases as many slabs as possible for a cache.
2300 * To help debugging, a zero exit status indicates all slabs were released.
2302 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2304 BUG_ON(!cachep
|| in_interrupt());
2306 return __cache_shrink(cachep
);
2308 EXPORT_SYMBOL(kmem_cache_shrink
);
2311 * kmem_cache_destroy - delete a cache
2312 * @cachep: the cache to destroy
2314 * Remove a struct kmem_cache object from the slab cache.
2315 * Returns 0 on success.
2317 * It is expected this function will be called by a module when it is
2318 * unloaded. This will remove the cache completely, and avoid a duplicate
2319 * cache being allocated each time a module is loaded and unloaded, if the
2320 * module doesn't have persistent in-kernel storage across loads and unloads.
2322 * The cache must be empty before calling this function.
2324 * The caller must guarantee that noone will allocate memory from the cache
2325 * during the kmem_cache_destroy().
2327 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2330 struct kmem_list3
*l3
;
2332 BUG_ON(!cachep
|| in_interrupt());
2334 /* Don't let CPUs to come and go */
2337 /* Find the cache in the chain of caches. */
2338 mutex_lock(&cache_chain_mutex
);
2340 * the chain is never empty, cache_cache is never destroyed
2342 list_del(&cachep
->next
);
2343 mutex_unlock(&cache_chain_mutex
);
2345 if (__cache_shrink(cachep
)) {
2346 slab_error(cachep
, "Can't free all objects");
2347 mutex_lock(&cache_chain_mutex
);
2348 list_add(&cachep
->next
, &cache_chain
);
2349 mutex_unlock(&cache_chain_mutex
);
2350 unlock_cpu_hotplug();
2354 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2357 for_each_online_cpu(i
)
2358 kfree(cachep
->array
[i
]);
2360 /* NUMA: free the list3 structures */
2361 for_each_online_node(i
) {
2362 l3
= cachep
->nodelists
[i
];
2365 free_alien_cache(l3
->alien
);
2369 kmem_cache_free(&cache_cache
, cachep
);
2370 unlock_cpu_hotplug();
2373 EXPORT_SYMBOL(kmem_cache_destroy
);
2375 /* Get the memory for a slab management obj. */
2376 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2377 int colour_off
, gfp_t local_flags
,
2382 if (OFF_SLAB(cachep
)) {
2383 /* Slab management obj is off-slab. */
2384 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2385 local_flags
, nodeid
);
2389 slabp
= objp
+ colour_off
;
2390 colour_off
+= cachep
->slab_size
;
2393 slabp
->colouroff
= colour_off
;
2394 slabp
->s_mem
= objp
+ colour_off
;
2395 slabp
->nodeid
= nodeid
;
2399 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2401 return (kmem_bufctl_t
*) (slabp
+ 1);
2404 static void cache_init_objs(struct kmem_cache
*cachep
,
2405 struct slab
*slabp
, unsigned long ctor_flags
)
2409 for (i
= 0; i
< cachep
->num
; i
++) {
2410 void *objp
= index_to_obj(cachep
, slabp
, i
);
2412 /* need to poison the objs? */
2413 if (cachep
->flags
& SLAB_POISON
)
2414 poison_obj(cachep
, objp
, POISON_FREE
);
2415 if (cachep
->flags
& SLAB_STORE_USER
)
2416 *dbg_userword(cachep
, objp
) = NULL
;
2418 if (cachep
->flags
& SLAB_RED_ZONE
) {
2419 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2420 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2423 * Constructors are not allowed to allocate memory from the same
2424 * cache which they are a constructor for. Otherwise, deadlock.
2425 * They must also be threaded.
2427 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2428 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2431 if (cachep
->flags
& SLAB_RED_ZONE
) {
2432 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2433 slab_error(cachep
, "constructor overwrote the"
2434 " end of an object");
2435 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2436 slab_error(cachep
, "constructor overwrote the"
2437 " start of an object");
2439 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2440 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2441 kernel_map_pages(virt_to_page(objp
),
2442 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2445 cachep
->ctor(objp
, cachep
, ctor_flags
);
2447 slab_bufctl(slabp
)[i
] = i
+ 1;
2449 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2453 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2455 if (flags
& SLAB_DMA
)
2456 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2458 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2461 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2464 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2468 next
= slab_bufctl(slabp
)[slabp
->free
];
2470 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2471 WARN_ON(slabp
->nodeid
!= nodeid
);
2478 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2479 void *objp
, int nodeid
)
2481 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2484 /* Verify that the slab belongs to the intended node */
2485 WARN_ON(slabp
->nodeid
!= nodeid
);
2487 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2488 printk(KERN_ERR
"slab: double free detected in cache "
2489 "'%s', objp %p\n", cachep
->name
, objp
);
2493 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2494 slabp
->free
= objnr
;
2499 * Map pages beginning at addr to the given cache and slab. This is required
2500 * for the slab allocator to be able to lookup the cache and slab of a
2501 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2503 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2509 page
= virt_to_page(addr
);
2512 if (likely(!PageCompound(page
)))
2513 nr_pages
<<= cache
->gfporder
;
2516 page_set_cache(page
, cache
);
2517 page_set_slab(page
, slab
);
2519 } while (--nr_pages
);
2523 * Grow (by 1) the number of slabs within a cache. This is called by
2524 * kmem_cache_alloc() when there are no active objs left in a cache.
2526 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2532 unsigned long ctor_flags
;
2533 struct kmem_list3
*l3
;
2536 * Be lazy and only check for valid flags here, keeping it out of the
2537 * critical path in kmem_cache_alloc().
2539 BUG_ON(flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
));
2540 if (flags
& SLAB_NO_GROW
)
2543 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2544 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2545 if (!(local_flags
& __GFP_WAIT
))
2547 * Not allowed to sleep. Need to tell a constructor about
2548 * this - it might need to know...
2550 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2552 /* Take the l3 list lock to change the colour_next on this node */
2554 l3
= cachep
->nodelists
[nodeid
];
2555 spin_lock(&l3
->list_lock
);
2557 /* Get colour for the slab, and cal the next value. */
2558 offset
= l3
->colour_next
;
2560 if (l3
->colour_next
>= cachep
->colour
)
2561 l3
->colour_next
= 0;
2562 spin_unlock(&l3
->list_lock
);
2564 offset
*= cachep
->colour_off
;
2566 if (local_flags
& __GFP_WAIT
)
2570 * The test for missing atomic flag is performed here, rather than
2571 * the more obvious place, simply to reduce the critical path length
2572 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2573 * will eventually be caught here (where it matters).
2575 kmem_flagcheck(cachep
, flags
);
2578 * Get mem for the objs. Attempt to allocate a physical page from
2581 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2585 /* Get slab management. */
2586 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
, nodeid
);
2590 slabp
->nodeid
= nodeid
;
2591 slab_map_pages(cachep
, slabp
, objp
);
2593 cache_init_objs(cachep
, slabp
, ctor_flags
);
2595 if (local_flags
& __GFP_WAIT
)
2596 local_irq_disable();
2598 spin_lock(&l3
->list_lock
);
2600 /* Make slab active. */
2601 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2602 STATS_INC_GROWN(cachep
);
2603 l3
->free_objects
+= cachep
->num
;
2604 spin_unlock(&l3
->list_lock
);
2607 kmem_freepages(cachep
, objp
);
2609 if (local_flags
& __GFP_WAIT
)
2610 local_irq_disable();
2617 * Perform extra freeing checks:
2618 * - detect bad pointers.
2619 * - POISON/RED_ZONE checking
2620 * - destructor calls, for caches with POISON+dtor
2622 static void kfree_debugcheck(const void *objp
)
2626 if (!virt_addr_valid(objp
)) {
2627 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2628 (unsigned long)objp
);
2631 page
= virt_to_page(objp
);
2632 if (!PageSlab(page
)) {
2633 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2634 (unsigned long)objp
);
2639 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2641 unsigned long redzone1
, redzone2
;
2643 redzone1
= *dbg_redzone1(cache
, obj
);
2644 redzone2
= *dbg_redzone2(cache
, obj
);
2649 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2652 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2653 slab_error(cache
, "double free detected");
2655 slab_error(cache
, "memory outside object was overwritten");
2657 printk(KERN_ERR
"%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2658 obj
, redzone1
, redzone2
);
2661 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2668 objp
-= obj_offset(cachep
);
2669 kfree_debugcheck(objp
);
2670 page
= virt_to_page(objp
);
2672 slabp
= page_get_slab(page
);
2674 if (cachep
->flags
& SLAB_RED_ZONE
) {
2675 verify_redzone_free(cachep
, objp
);
2676 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2677 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2679 if (cachep
->flags
& SLAB_STORE_USER
)
2680 *dbg_userword(cachep
, objp
) = caller
;
2682 objnr
= obj_to_index(cachep
, slabp
, objp
);
2684 BUG_ON(objnr
>= cachep
->num
);
2685 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2687 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2689 * Need to call the slab's constructor so the caller can
2690 * perform a verify of its state (debugging). Called without
2691 * the cache-lock held.
2693 cachep
->ctor(objp
+ obj_offset(cachep
),
2694 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2696 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2697 /* we want to cache poison the object,
2698 * call the destruction callback
2700 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2702 #ifdef CONFIG_DEBUG_SLAB_LEAK
2703 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2705 if (cachep
->flags
& SLAB_POISON
) {
2706 #ifdef CONFIG_DEBUG_PAGEALLOC
2707 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2708 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2709 kernel_map_pages(virt_to_page(objp
),
2710 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2712 poison_obj(cachep
, objp
, POISON_FREE
);
2715 poison_obj(cachep
, objp
, POISON_FREE
);
2721 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2726 /* Check slab's freelist to see if this obj is there. */
2727 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2729 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2732 if (entries
!= cachep
->num
- slabp
->inuse
) {
2734 printk(KERN_ERR
"slab: Internal list corruption detected in "
2735 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2736 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2738 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2741 printk("\n%03x:", i
);
2742 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2749 #define kfree_debugcheck(x) do { } while(0)
2750 #define cache_free_debugcheck(x,objp,z) (objp)
2751 #define check_slabp(x,y) do { } while(0)
2754 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2757 struct kmem_list3
*l3
;
2758 struct array_cache
*ac
;
2761 ac
= cpu_cache_get(cachep
);
2763 batchcount
= ac
->batchcount
;
2764 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2766 * If there was little recent activity on this cache, then
2767 * perform only a partial refill. Otherwise we could generate
2770 batchcount
= BATCHREFILL_LIMIT
;
2772 l3
= cachep
->nodelists
[numa_node_id()];
2774 BUG_ON(ac
->avail
> 0 || !l3
);
2775 spin_lock(&l3
->list_lock
);
2777 /* See if we can refill from the shared array */
2778 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2781 while (batchcount
> 0) {
2782 struct list_head
*entry
;
2784 /* Get slab alloc is to come from. */
2785 entry
= l3
->slabs_partial
.next
;
2786 if (entry
== &l3
->slabs_partial
) {
2787 l3
->free_touched
= 1;
2788 entry
= l3
->slabs_free
.next
;
2789 if (entry
== &l3
->slabs_free
)
2793 slabp
= list_entry(entry
, struct slab
, list
);
2794 check_slabp(cachep
, slabp
);
2795 check_spinlock_acquired(cachep
);
2796 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2797 STATS_INC_ALLOCED(cachep
);
2798 STATS_INC_ACTIVE(cachep
);
2799 STATS_SET_HIGH(cachep
);
2801 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2804 check_slabp(cachep
, slabp
);
2806 /* move slabp to correct slabp list: */
2807 list_del(&slabp
->list
);
2808 if (slabp
->free
== BUFCTL_END
)
2809 list_add(&slabp
->list
, &l3
->slabs_full
);
2811 list_add(&slabp
->list
, &l3
->slabs_partial
);
2815 l3
->free_objects
-= ac
->avail
;
2817 spin_unlock(&l3
->list_lock
);
2819 if (unlikely(!ac
->avail
)) {
2821 x
= cache_grow(cachep
, flags
, numa_node_id());
2823 /* cache_grow can reenable interrupts, then ac could change. */
2824 ac
= cpu_cache_get(cachep
);
2825 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2828 if (!ac
->avail
) /* objects refilled by interrupt? */
2832 return ac
->entry
[--ac
->avail
];
2835 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2838 might_sleep_if(flags
& __GFP_WAIT
);
2840 kmem_flagcheck(cachep
, flags
);
2845 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2846 gfp_t flags
, void *objp
, void *caller
)
2850 if (cachep
->flags
& SLAB_POISON
) {
2851 #ifdef CONFIG_DEBUG_PAGEALLOC
2852 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2853 kernel_map_pages(virt_to_page(objp
),
2854 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2856 check_poison_obj(cachep
, objp
);
2858 check_poison_obj(cachep
, objp
);
2860 poison_obj(cachep
, objp
, POISON_INUSE
);
2862 if (cachep
->flags
& SLAB_STORE_USER
)
2863 *dbg_userword(cachep
, objp
) = caller
;
2865 if (cachep
->flags
& SLAB_RED_ZONE
) {
2866 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2867 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2868 slab_error(cachep
, "double free, or memory outside"
2869 " object was overwritten");
2871 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2872 objp
, *dbg_redzone1(cachep
, objp
),
2873 *dbg_redzone2(cachep
, objp
));
2875 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2876 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2878 #ifdef CONFIG_DEBUG_SLAB_LEAK
2883 slabp
= page_get_slab(virt_to_page(objp
));
2884 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2885 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
2888 objp
+= obj_offset(cachep
);
2889 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2890 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2892 if (!(flags
& __GFP_WAIT
))
2893 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2895 cachep
->ctor(objp
, cachep
, ctor_flags
);
2900 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2903 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2906 struct array_cache
*ac
;
2909 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
2910 objp
= alternate_node_alloc(cachep
, flags
);
2917 ac
= cpu_cache_get(cachep
);
2918 if (likely(ac
->avail
)) {
2919 STATS_INC_ALLOCHIT(cachep
);
2921 objp
= ac
->entry
[--ac
->avail
];
2923 STATS_INC_ALLOCMISS(cachep
);
2924 objp
= cache_alloc_refill(cachep
, flags
);
2929 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
2930 gfp_t flags
, void *caller
)
2932 unsigned long save_flags
;
2935 cache_alloc_debugcheck_before(cachep
, flags
);
2937 local_irq_save(save_flags
);
2938 objp
= ____cache_alloc(cachep
, flags
);
2939 local_irq_restore(save_flags
);
2940 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2948 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2950 * If we are in_interrupt, then process context, including cpusets and
2951 * mempolicy, may not apply and should not be used for allocation policy.
2953 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2955 int nid_alloc
, nid_here
;
2959 nid_alloc
= nid_here
= numa_node_id();
2960 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2961 nid_alloc
= cpuset_mem_spread_node();
2962 else if (current
->mempolicy
)
2963 nid_alloc
= slab_node(current
->mempolicy
);
2964 if (nid_alloc
!= nid_here
)
2965 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
2970 * A interface to enable slab creation on nodeid
2972 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
2975 struct list_head
*entry
;
2977 struct kmem_list3
*l3
;
2981 l3
= cachep
->nodelists
[nodeid
];
2986 spin_lock(&l3
->list_lock
);
2987 entry
= l3
->slabs_partial
.next
;
2988 if (entry
== &l3
->slabs_partial
) {
2989 l3
->free_touched
= 1;
2990 entry
= l3
->slabs_free
.next
;
2991 if (entry
== &l3
->slabs_free
)
2995 slabp
= list_entry(entry
, struct slab
, list
);
2996 check_spinlock_acquired_node(cachep
, nodeid
);
2997 check_slabp(cachep
, slabp
);
2999 STATS_INC_NODEALLOCS(cachep
);
3000 STATS_INC_ACTIVE(cachep
);
3001 STATS_SET_HIGH(cachep
);
3003 BUG_ON(slabp
->inuse
== cachep
->num
);
3005 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3006 check_slabp(cachep
, slabp
);
3008 /* move slabp to correct slabp list: */
3009 list_del(&slabp
->list
);
3011 if (slabp
->free
== BUFCTL_END
)
3012 list_add(&slabp
->list
, &l3
->slabs_full
);
3014 list_add(&slabp
->list
, &l3
->slabs_partial
);
3016 spin_unlock(&l3
->list_lock
);
3020 spin_unlock(&l3
->list_lock
);
3021 x
= cache_grow(cachep
, flags
, nodeid
);
3033 * Caller needs to acquire correct kmem_list's list_lock
3035 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3039 struct kmem_list3
*l3
;
3041 for (i
= 0; i
< nr_objects
; i
++) {
3042 void *objp
= objpp
[i
];
3045 slabp
= virt_to_slab(objp
);
3046 l3
= cachep
->nodelists
[node
];
3047 list_del(&slabp
->list
);
3048 check_spinlock_acquired_node(cachep
, node
);
3049 check_slabp(cachep
, slabp
);
3050 slab_put_obj(cachep
, slabp
, objp
, node
);
3051 STATS_DEC_ACTIVE(cachep
);
3053 check_slabp(cachep
, slabp
);
3055 /* fixup slab chains */
3056 if (slabp
->inuse
== 0) {
3057 if (l3
->free_objects
> l3
->free_limit
) {
3058 l3
->free_objects
-= cachep
->num
;
3059 slab_destroy(cachep
, slabp
);
3061 list_add(&slabp
->list
, &l3
->slabs_free
);
3064 /* Unconditionally move a slab to the end of the
3065 * partial list on free - maximum time for the
3066 * other objects to be freed, too.
3068 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3073 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3076 struct kmem_list3
*l3
;
3077 int node
= numa_node_id();
3079 batchcount
= ac
->batchcount
;
3081 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3084 l3
= cachep
->nodelists
[node
];
3085 spin_lock(&l3
->list_lock
);
3087 struct array_cache
*shared_array
= l3
->shared
;
3088 int max
= shared_array
->limit
- shared_array
->avail
;
3090 if (batchcount
> max
)
3092 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3093 ac
->entry
, sizeof(void *) * batchcount
);
3094 shared_array
->avail
+= batchcount
;
3099 free_block(cachep
, ac
->entry
, batchcount
, node
);
3104 struct list_head
*p
;
3106 p
= l3
->slabs_free
.next
;
3107 while (p
!= &(l3
->slabs_free
)) {
3110 slabp
= list_entry(p
, struct slab
, list
);
3111 BUG_ON(slabp
->inuse
);
3116 STATS_SET_FREEABLE(cachep
, i
);
3119 spin_unlock(&l3
->list_lock
);
3120 ac
->avail
-= batchcount
;
3121 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3125 * Release an obj back to its cache. If the obj has a constructed state, it must
3126 * be in this state _before_ it is released. Called with disabled ints.
3128 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3130 struct array_cache
*ac
= cpu_cache_get(cachep
);
3133 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3135 if (cache_free_alien(cachep
, objp
))
3138 if (likely(ac
->avail
< ac
->limit
)) {
3139 STATS_INC_FREEHIT(cachep
);
3140 ac
->entry
[ac
->avail
++] = objp
;
3143 STATS_INC_FREEMISS(cachep
);
3144 cache_flusharray(cachep
, ac
);
3145 ac
->entry
[ac
->avail
++] = objp
;
3150 * kmem_cache_alloc - Allocate an object
3151 * @cachep: The cache to allocate from.
3152 * @flags: See kmalloc().
3154 * Allocate an object from this cache. The flags are only relevant
3155 * if the cache has no available objects.
3157 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3159 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3161 EXPORT_SYMBOL(kmem_cache_alloc
);
3164 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3165 * @cache: The cache to allocate from.
3166 * @flags: See kmalloc().
3168 * Allocate an object from this cache and set the allocated memory to zero.
3169 * The flags are only relevant if the cache has no available objects.
3171 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3173 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3175 memset(ret
, 0, obj_size(cache
));
3178 EXPORT_SYMBOL(kmem_cache_zalloc
);
3181 * kmem_ptr_validate - check if an untrusted pointer might
3183 * @cachep: the cache we're checking against
3184 * @ptr: pointer to validate
3186 * This verifies that the untrusted pointer looks sane:
3187 * it is _not_ a guarantee that the pointer is actually
3188 * part of the slab cache in question, but it at least
3189 * validates that the pointer can be dereferenced and
3190 * looks half-way sane.
3192 * Currently only used for dentry validation.
3194 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3196 unsigned long addr
= (unsigned long)ptr
;
3197 unsigned long min_addr
= PAGE_OFFSET
;
3198 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3199 unsigned long size
= cachep
->buffer_size
;
3202 if (unlikely(addr
< min_addr
))
3204 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3206 if (unlikely(addr
& align_mask
))
3208 if (unlikely(!kern_addr_valid(addr
)))
3210 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3212 page
= virt_to_page(ptr
);
3213 if (unlikely(!PageSlab(page
)))
3215 if (unlikely(page_get_cache(page
) != cachep
))
3224 * kmem_cache_alloc_node - Allocate an object on the specified node
3225 * @cachep: The cache to allocate from.
3226 * @flags: See kmalloc().
3227 * @nodeid: node number of the target node.
3229 * Identical to kmem_cache_alloc, except that this function is slow
3230 * and can sleep. And it will allocate memory on the given node, which
3231 * can improve the performance for cpu bound structures.
3232 * New and improved: it will now make sure that the object gets
3233 * put on the correct node list so that there is no false sharing.
3235 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3237 unsigned long save_flags
;
3240 cache_alloc_debugcheck_before(cachep
, flags
);
3241 local_irq_save(save_flags
);
3243 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3244 !cachep
->nodelists
[nodeid
])
3245 ptr
= ____cache_alloc(cachep
, flags
);
3247 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3248 local_irq_restore(save_flags
);
3250 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3251 __builtin_return_address(0));
3255 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3257 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3259 struct kmem_cache
*cachep
;
3261 cachep
= kmem_find_general_cachep(size
, flags
);
3262 if (unlikely(cachep
== NULL
))
3264 return kmem_cache_alloc_node(cachep
, flags
, node
);
3266 EXPORT_SYMBOL(kmalloc_node
);
3270 * __do_kmalloc - allocate memory
3271 * @size: how many bytes of memory are required.
3272 * @flags: the type of memory to allocate (see kmalloc).
3273 * @caller: function caller for debug tracking of the caller
3275 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3278 struct kmem_cache
*cachep
;
3280 /* If you want to save a few bytes .text space: replace
3282 * Then kmalloc uses the uninlined functions instead of the inline
3285 cachep
= __find_general_cachep(size
, flags
);
3286 if (unlikely(cachep
== NULL
))
3288 return __cache_alloc(cachep
, flags
, caller
);
3292 void *__kmalloc(size_t size
, gfp_t flags
)
3294 #ifndef CONFIG_DEBUG_SLAB
3295 return __do_kmalloc(size
, flags
, NULL
);
3297 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3300 EXPORT_SYMBOL(__kmalloc
);
3302 #ifdef CONFIG_DEBUG_SLAB
3303 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3305 return __do_kmalloc(size
, flags
, caller
);
3307 EXPORT_SYMBOL(__kmalloc_track_caller
);
3312 * __alloc_percpu - allocate one copy of the object for every present
3313 * cpu in the system, zeroing them.
3314 * Objects should be dereferenced using the per_cpu_ptr macro only.
3316 * @size: how many bytes of memory are required.
3318 void *__alloc_percpu(size_t size
)
3321 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3327 * Cannot use for_each_online_cpu since a cpu may come online
3328 * and we have no way of figuring out how to fix the array
3329 * that we have allocated then....
3331 for_each_possible_cpu(i
) {
3332 int node
= cpu_to_node(i
);
3334 if (node_online(node
))
3335 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3337 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3339 if (!pdata
->ptrs
[i
])
3341 memset(pdata
->ptrs
[i
], 0, size
);
3344 /* Catch derefs w/o wrappers */
3345 return (void *)(~(unsigned long)pdata
);
3349 if (!cpu_possible(i
))
3351 kfree(pdata
->ptrs
[i
]);
3356 EXPORT_SYMBOL(__alloc_percpu
);
3360 * kmem_cache_free - Deallocate an object
3361 * @cachep: The cache the allocation was from.
3362 * @objp: The previously allocated object.
3364 * Free an object which was previously allocated from this
3367 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3369 unsigned long flags
;
3371 BUG_ON(virt_to_cache(objp
) != cachep
);
3373 local_irq_save(flags
);
3374 __cache_free(cachep
, objp
);
3375 local_irq_restore(flags
);
3377 EXPORT_SYMBOL(kmem_cache_free
);
3380 * kfree - free previously allocated memory
3381 * @objp: pointer returned by kmalloc.
3383 * If @objp is NULL, no operation is performed.
3385 * Don't free memory not originally allocated by kmalloc()
3386 * or you will run into trouble.
3388 void kfree(const void *objp
)
3390 struct kmem_cache
*c
;
3391 unsigned long flags
;
3393 if (unlikely(!objp
))
3395 local_irq_save(flags
);
3396 kfree_debugcheck(objp
);
3397 c
= virt_to_cache(objp
);
3398 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3399 __cache_free(c
, (void *)objp
);
3400 local_irq_restore(flags
);
3402 EXPORT_SYMBOL(kfree
);
3406 * free_percpu - free previously allocated percpu memory
3407 * @objp: pointer returned by alloc_percpu.
3409 * Don't free memory not originally allocated by alloc_percpu()
3410 * The complemented objp is to check for that.
3412 void free_percpu(const void *objp
)
3415 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3418 * We allocate for all cpus so we cannot use for online cpu here.
3420 for_each_possible_cpu(i
)
3424 EXPORT_SYMBOL(free_percpu
);
3427 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3429 return obj_size(cachep
);
3431 EXPORT_SYMBOL(kmem_cache_size
);
3433 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3435 return cachep
->name
;
3437 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3440 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3442 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3445 struct kmem_list3
*l3
;
3446 struct array_cache
*new_shared
;
3447 struct array_cache
**new_alien
;
3449 for_each_online_node(node
) {
3451 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3455 new_shared
= alloc_arraycache(node
,
3456 cachep
->shared
*cachep
->batchcount
,
3459 free_alien_cache(new_alien
);
3463 l3
= cachep
->nodelists
[node
];
3465 struct array_cache
*shared
= l3
->shared
;
3467 spin_lock_irq(&l3
->list_lock
);
3470 free_block(cachep
, shared
->entry
,
3471 shared
->avail
, node
);
3473 l3
->shared
= new_shared
;
3475 l3
->alien
= new_alien
;
3478 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3479 cachep
->batchcount
+ cachep
->num
;
3480 spin_unlock_irq(&l3
->list_lock
);
3482 free_alien_cache(new_alien
);
3485 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3487 free_alien_cache(new_alien
);
3492 kmem_list3_init(l3
);
3493 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3494 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3495 l3
->shared
= new_shared
;
3496 l3
->alien
= new_alien
;
3497 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3498 cachep
->batchcount
+ cachep
->num
;
3499 cachep
->nodelists
[node
] = l3
;
3504 if (!cachep
->next
.next
) {
3505 /* Cache is not active yet. Roll back what we did */
3508 if (cachep
->nodelists
[node
]) {
3509 l3
= cachep
->nodelists
[node
];
3512 free_alien_cache(l3
->alien
);
3514 cachep
->nodelists
[node
] = NULL
;
3522 struct ccupdate_struct
{
3523 struct kmem_cache
*cachep
;
3524 struct array_cache
*new[NR_CPUS
];
3527 static void do_ccupdate_local(void *info
)
3529 struct ccupdate_struct
*new = info
;
3530 struct array_cache
*old
;
3533 old
= cpu_cache_get(new->cachep
);
3535 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3536 new->new[smp_processor_id()] = old
;
3539 /* Always called with the cache_chain_mutex held */
3540 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3541 int batchcount
, int shared
)
3543 struct ccupdate_struct
new;
3546 memset(&new.new, 0, sizeof(new.new));
3547 for_each_online_cpu(i
) {
3548 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3551 for (i
--; i
>= 0; i
--)
3556 new.cachep
= cachep
;
3558 on_each_cpu(do_ccupdate_local
, (void *)&new, 1, 1);
3561 cachep
->batchcount
= batchcount
;
3562 cachep
->limit
= limit
;
3563 cachep
->shared
= shared
;
3565 for_each_online_cpu(i
) {
3566 struct array_cache
*ccold
= new.new[i
];
3569 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3570 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3571 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3575 err
= alloc_kmemlist(cachep
);
3577 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3578 cachep
->name
, -err
);
3584 /* Called with cache_chain_mutex held always */
3585 static void enable_cpucache(struct kmem_cache
*cachep
)
3591 * The head array serves three purposes:
3592 * - create a LIFO ordering, i.e. return objects that are cache-warm
3593 * - reduce the number of spinlock operations.
3594 * - reduce the number of linked list operations on the slab and
3595 * bufctl chains: array operations are cheaper.
3596 * The numbers are guessed, we should auto-tune as described by
3599 if (cachep
->buffer_size
> 131072)
3601 else if (cachep
->buffer_size
> PAGE_SIZE
)
3603 else if (cachep
->buffer_size
> 1024)
3605 else if (cachep
->buffer_size
> 256)
3611 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3612 * allocation behaviour: Most allocs on one cpu, most free operations
3613 * on another cpu. For these cases, an efficient object passing between
3614 * cpus is necessary. This is provided by a shared array. The array
3615 * replaces Bonwick's magazine layer.
3616 * On uniprocessor, it's functionally equivalent (but less efficient)
3617 * to a larger limit. Thus disabled by default.
3621 if (cachep
->buffer_size
<= PAGE_SIZE
)
3627 * With debugging enabled, large batchcount lead to excessively long
3628 * periods with disabled local interrupts. Limit the batchcount
3633 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3635 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3636 cachep
->name
, -err
);
3640 * Drain an array if it contains any elements taking the l3 lock only if
3641 * necessary. Note that the l3 listlock also protects the array_cache
3642 * if drain_array() is used on the shared array.
3644 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3645 struct array_cache
*ac
, int force
, int node
)
3649 if (!ac
|| !ac
->avail
)
3651 if (ac
->touched
&& !force
) {
3654 spin_lock_irq(&l3
->list_lock
);
3656 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3657 if (tofree
> ac
->avail
)
3658 tofree
= (ac
->avail
+ 1) / 2;
3659 free_block(cachep
, ac
->entry
, tofree
, node
);
3660 ac
->avail
-= tofree
;
3661 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3662 sizeof(void *) * ac
->avail
);
3664 spin_unlock_irq(&l3
->list_lock
);
3669 * cache_reap - Reclaim memory from caches.
3670 * @unused: unused parameter
3672 * Called from workqueue/eventd every few seconds.
3674 * - clear the per-cpu caches for this CPU.
3675 * - return freeable pages to the main free memory pool.
3677 * If we cannot acquire the cache chain mutex then just give up - we'll try
3678 * again on the next iteration.
3680 static void cache_reap(void *unused
)
3682 struct kmem_cache
*searchp
;
3683 struct kmem_list3
*l3
;
3684 int node
= numa_node_id();
3686 if (!mutex_trylock(&cache_chain_mutex
)) {
3687 /* Give up. Setup the next iteration. */
3688 schedule_delayed_work(&__get_cpu_var(reap_work
),
3693 list_for_each_entry(searchp
, &cache_chain
, next
) {
3694 struct list_head
*p
;
3701 * We only take the l3 lock if absolutely necessary and we
3702 * have established with reasonable certainty that
3703 * we can do some work if the lock was obtained.
3705 l3
= searchp
->nodelists
[node
];
3707 reap_alien(searchp
, l3
);
3709 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3712 * These are racy checks but it does not matter
3713 * if we skip one check or scan twice.
3715 if (time_after(l3
->next_reap
, jiffies
))
3718 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3720 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3722 if (l3
->free_touched
) {
3723 l3
->free_touched
= 0;
3727 tofree
= (l3
->free_limit
+ 5 * searchp
->num
- 1) /
3731 * Do not lock if there are no free blocks.
3733 if (list_empty(&l3
->slabs_free
))
3736 spin_lock_irq(&l3
->list_lock
);
3737 p
= l3
->slabs_free
.next
;
3738 if (p
== &(l3
->slabs_free
)) {
3739 spin_unlock_irq(&l3
->list_lock
);
3743 slabp
= list_entry(p
, struct slab
, list
);
3744 BUG_ON(slabp
->inuse
);
3745 list_del(&slabp
->list
);
3746 STATS_INC_REAPED(searchp
);
3749 * Safe to drop the lock. The slab is no longer linked
3750 * to the cache. searchp cannot disappear, we hold
3753 l3
->free_objects
-= searchp
->num
;
3754 spin_unlock_irq(&l3
->list_lock
);
3755 slab_destroy(searchp
, slabp
);
3756 } while (--tofree
> 0);
3761 mutex_unlock(&cache_chain_mutex
);
3763 /* Set up the next iteration */
3764 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3767 #ifdef CONFIG_PROC_FS
3769 static void print_slabinfo_header(struct seq_file
*m
)
3772 * Output format version, so at least we can change it
3773 * without _too_ many complaints.
3776 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3778 seq_puts(m
, "slabinfo - version: 2.1\n");
3780 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3781 "<objperslab> <pagesperslab>");
3782 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3783 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3785 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3786 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3787 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3792 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3795 struct list_head
*p
;
3797 mutex_lock(&cache_chain_mutex
);
3799 print_slabinfo_header(m
);
3800 p
= cache_chain
.next
;
3803 if (p
== &cache_chain
)
3806 return list_entry(p
, struct kmem_cache
, next
);
3809 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3811 struct kmem_cache
*cachep
= p
;
3813 return cachep
->next
.next
== &cache_chain
?
3814 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3817 static void s_stop(struct seq_file
*m
, void *p
)
3819 mutex_unlock(&cache_chain_mutex
);
3822 static int s_show(struct seq_file
*m
, void *p
)
3824 struct kmem_cache
*cachep
= p
;
3826 unsigned long active_objs
;
3827 unsigned long num_objs
;
3828 unsigned long active_slabs
= 0;
3829 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3833 struct kmem_list3
*l3
;
3837 for_each_online_node(node
) {
3838 l3
= cachep
->nodelists
[node
];
3843 spin_lock_irq(&l3
->list_lock
);
3845 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
3846 if (slabp
->inuse
!= cachep
->num
&& !error
)
3847 error
= "slabs_full accounting error";
3848 active_objs
+= cachep
->num
;
3851 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
3852 if (slabp
->inuse
== cachep
->num
&& !error
)
3853 error
= "slabs_partial inuse accounting error";
3854 if (!slabp
->inuse
&& !error
)
3855 error
= "slabs_partial/inuse accounting error";
3856 active_objs
+= slabp
->inuse
;
3859 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
3860 if (slabp
->inuse
&& !error
)
3861 error
= "slabs_free/inuse accounting error";
3864 free_objects
+= l3
->free_objects
;
3866 shared_avail
+= l3
->shared
->avail
;
3868 spin_unlock_irq(&l3
->list_lock
);
3870 num_slabs
+= active_slabs
;
3871 num_objs
= num_slabs
* cachep
->num
;
3872 if (num_objs
- active_objs
!= free_objects
&& !error
)
3873 error
= "free_objects accounting error";
3875 name
= cachep
->name
;
3877 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3879 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3880 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3881 cachep
->num
, (1 << cachep
->gfporder
));
3882 seq_printf(m
, " : tunables %4u %4u %4u",
3883 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3884 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3885 active_slabs
, num_slabs
, shared_avail
);
3888 unsigned long high
= cachep
->high_mark
;
3889 unsigned long allocs
= cachep
->num_allocations
;
3890 unsigned long grown
= cachep
->grown
;
3891 unsigned long reaped
= cachep
->reaped
;
3892 unsigned long errors
= cachep
->errors
;
3893 unsigned long max_freeable
= cachep
->max_freeable
;
3894 unsigned long node_allocs
= cachep
->node_allocs
;
3895 unsigned long node_frees
= cachep
->node_frees
;
3896 unsigned long overflows
= cachep
->node_overflow
;
3898 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3899 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3900 reaped
, errors
, max_freeable
, node_allocs
,
3901 node_frees
, overflows
);
3905 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3906 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3907 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3908 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3910 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3911 allochit
, allocmiss
, freehit
, freemiss
);
3919 * slabinfo_op - iterator that generates /proc/slabinfo
3928 * num-pages-per-slab
3929 * + further values on SMP and with statistics enabled
3932 struct seq_operations slabinfo_op
= {
3939 #define MAX_SLABINFO_WRITE 128
3941 * slabinfo_write - Tuning for the slab allocator
3943 * @buffer: user buffer
3944 * @count: data length
3947 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3948 size_t count
, loff_t
*ppos
)
3950 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3951 int limit
, batchcount
, shared
, res
;
3952 struct kmem_cache
*cachep
;
3954 if (count
> MAX_SLABINFO_WRITE
)
3956 if (copy_from_user(&kbuf
, buffer
, count
))
3958 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3960 tmp
= strchr(kbuf
, ' ');
3965 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3968 /* Find the cache in the chain of caches. */
3969 mutex_lock(&cache_chain_mutex
);
3971 list_for_each_entry(cachep
, &cache_chain
, next
) {
3972 if (!strcmp(cachep
->name
, kbuf
)) {
3973 if (limit
< 1 || batchcount
< 1 ||
3974 batchcount
> limit
|| shared
< 0) {
3977 res
= do_tune_cpucache(cachep
, limit
,
3978 batchcount
, shared
);
3983 mutex_unlock(&cache_chain_mutex
);
3989 #ifdef CONFIG_DEBUG_SLAB_LEAK
3991 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
3994 struct list_head
*p
;
3996 mutex_lock(&cache_chain_mutex
);
3997 p
= cache_chain
.next
;
4000 if (p
== &cache_chain
)
4003 return list_entry(p
, struct kmem_cache
, next
);
4006 static inline int add_caller(unsigned long *n
, unsigned long v
)
4016 unsigned long *q
= p
+ 2 * i
;
4030 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4036 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4042 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4043 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4045 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4050 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4052 #ifdef CONFIG_KALLSYMS
4055 unsigned long offset
, size
;
4056 char namebuf
[KSYM_NAME_LEN
+1];
4058 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4061 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4063 seq_printf(m
, " [%s]", modname
);
4067 seq_printf(m
, "%p", (void *)address
);
4070 static int leaks_show(struct seq_file
*m
, void *p
)
4072 struct kmem_cache
*cachep
= p
;
4074 struct kmem_list3
*l3
;
4076 unsigned long *n
= m
->private;
4080 if (!(cachep
->flags
& SLAB_STORE_USER
))
4082 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4085 /* OK, we can do it */
4089 for_each_online_node(node
) {
4090 l3
= cachep
->nodelists
[node
];
4095 spin_lock_irq(&l3
->list_lock
);
4097 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4098 handle_slab(n
, cachep
, slabp
);
4099 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4100 handle_slab(n
, cachep
, slabp
);
4101 spin_unlock_irq(&l3
->list_lock
);
4103 name
= cachep
->name
;
4105 /* Increase the buffer size */
4106 mutex_unlock(&cache_chain_mutex
);
4107 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4109 /* Too bad, we are really out */
4111 mutex_lock(&cache_chain_mutex
);
4114 *(unsigned long *)m
->private = n
[0] * 2;
4116 mutex_lock(&cache_chain_mutex
);
4117 /* Now make sure this entry will be retried */
4121 for (i
= 0; i
< n
[1]; i
++) {
4122 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4123 show_symbol(m
, n
[2*i
+2]);
4129 struct seq_operations slabstats_op
= {
4130 .start
= leaks_start
,
4139 * ksize - get the actual amount of memory allocated for a given object
4140 * @objp: Pointer to the object
4142 * kmalloc may internally round up allocations and return more memory
4143 * than requested. ksize() can be used to determine the actual amount of
4144 * memory allocated. The caller may use this additional memory, even though
4145 * a smaller amount of memory was initially specified with the kmalloc call.
4146 * The caller must guarantee that objp points to a valid object previously
4147 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4148 * must not be freed during the duration of the call.
4150 unsigned int ksize(const void *objp
)
4152 if (unlikely(objp
== NULL
))
4155 return obj_size(virt_to_cache(objp
));