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>
110 #include <linux/rtmutex.h>
112 #include <asm/uaccess.h>
113 #include <asm/cacheflush.h>
114 #include <asm/tlbflush.h>
115 #include <asm/page.h>
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
154 #define ARCH_KMALLOC_MINALIGN 0
157 #ifndef ARCH_SLAB_MINALIGN
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
165 #define ARCH_SLAB_MINALIGN 0
168 #ifndef ARCH_KMALLOC_FLAGS
169 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 /* Legal flag mask for kmem_cache_create(). */
174 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
181 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
190 * Bufctl's are used for linking objs within a slab
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 typedef unsigned int kmem_bufctl_t
;
207 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
209 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list
;
221 unsigned long colouroff
;
222 void *s_mem
; /* including colour offset */
223 unsigned int inuse
; /* num of objs active in slab */
225 unsigned short nodeid
;
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
242 * We assume struct slab_rcu can overlay struct slab when destroying.
245 struct rcu_head head
;
246 struct kmem_cache
*cachep
;
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
258 * The limit is stored in the per-cpu structure to reduce the data cache
265 unsigned int batchcount
;
266 unsigned int touched
;
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
272 * [0] is for gcc 2.95. It should really be [].
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init
{
282 struct array_cache cache
;
283 void *entries
[BOOT_CPUCACHE_ENTRIES
];
287 * The slab lists for all objects.
290 struct list_head slabs_partial
; /* partial list first, better asm code */
291 struct list_head slabs_full
;
292 struct list_head slabs_free
;
293 unsigned long free_objects
;
294 unsigned int free_limit
;
295 unsigned int colour_next
; /* Per-node cache coloring */
296 spinlock_t list_lock
;
297 struct array_cache
*shared
; /* shared per node */
298 struct array_cache
**alien
; /* on other nodes */
299 unsigned long next_reap
; /* updated without locking */
300 int free_touched
; /* updated without locking */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
312 static int drain_freelist(struct kmem_cache
*cache
,
313 struct kmem_list3
*l3
, int tofree
);
314 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
316 static void enable_cpucache(struct kmem_cache
*cachep
);
317 static void cache_reap(void *unused
);
320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
323 static __always_inline
int index_of(const size_t size
)
325 extern void __bad_size(void);
327 if (__builtin_constant_p(size
)) {
335 #include "linux/kmalloc_sizes.h"
343 static int slab_early_init
= 1;
345 #define INDEX_AC index_of(sizeof(struct arraycache_init))
346 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
348 static void kmem_list3_init(struct kmem_list3
*parent
)
350 INIT_LIST_HEAD(&parent
->slabs_full
);
351 INIT_LIST_HEAD(&parent
->slabs_partial
);
352 INIT_LIST_HEAD(&parent
->slabs_free
);
353 parent
->shared
= NULL
;
354 parent
->alien
= NULL
;
355 parent
->colour_next
= 0;
356 spin_lock_init(&parent
->list_lock
);
357 parent
->free_objects
= 0;
358 parent
->free_touched
= 0;
361 #define MAKE_LIST(cachep, listp, slab, nodeid) \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
367 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 /* 1) per-cpu data, touched during every alloc/free */
382 struct array_cache
*array
[NR_CPUS
];
383 /* 2) Cache tunables. Protected by cache_chain_mutex */
384 unsigned int batchcount
;
388 unsigned int buffer_size
;
389 /* 3) touched by every alloc & free from the backend */
390 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
392 unsigned int flags
; /* constant flags */
393 unsigned int num
; /* # of objs per slab */
395 /* 4) cache_grow/shrink */
396 /* order of pgs per slab (2^n) */
397 unsigned int gfporder
;
399 /* force GFP flags, e.g. GFP_DMA */
402 size_t colour
; /* cache colouring range */
403 unsigned int colour_off
; /* colour offset */
404 struct kmem_cache
*slabp_cache
;
405 unsigned int slab_size
;
406 unsigned int dflags
; /* dynamic flags */
408 /* constructor func */
409 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
411 /* de-constructor func */
412 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
414 /* 5) cache creation/removal */
416 struct list_head next
;
420 unsigned long num_active
;
421 unsigned long num_allocations
;
422 unsigned long high_mark
;
424 unsigned long reaped
;
425 unsigned long errors
;
426 unsigned long max_freeable
;
427 unsigned long node_allocs
;
428 unsigned long node_frees
;
429 unsigned long node_overflow
;
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
447 #define CFLGS_OFF_SLAB (0x80000000UL)
448 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
450 #define BATCHREFILL_LIMIT 16
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
455 * OTOH the cpuarrays can contain lots of objects,
456 * which could lock up otherwise freeable slabs.
458 #define REAPTIMEOUT_CPUC (2*HZ)
459 #define REAPTIMEOUT_LIST3 (4*HZ)
462 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
463 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465 #define STATS_INC_GROWN(x) ((x)->grown++)
466 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
467 #define STATS_SET_HIGH(x) \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
472 #define STATS_INC_ERR(x) ((x)->errors++)
473 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
474 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
475 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
476 #define STATS_SET_FREEABLE(x, i) \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
481 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
486 #define STATS_INC_ACTIVE(x) do { } while (0)
487 #define STATS_DEC_ACTIVE(x) do { } while (0)
488 #define STATS_INC_ALLOCED(x) do { } while (0)
489 #define STATS_INC_GROWN(x) do { } while (0)
490 #define STATS_ADD_REAPED(x,y) do { } while (0)
491 #define STATS_SET_HIGH(x) do { } while (0)
492 #define STATS_INC_ERR(x) do { } while (0)
493 #define STATS_INC_NODEALLOCS(x) do { } while (0)
494 #define STATS_INC_NODEFREES(x) do { } while (0)
495 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
496 #define STATS_SET_FREEABLE(x, i) do { } while (0)
497 #define STATS_INC_ALLOCHIT(x) do { } while (0)
498 #define STATS_INC_ALLOCMISS(x) do { } while (0)
499 #define STATS_INC_FREEHIT(x) do { } while (0)
500 #define STATS_INC_FREEMISS(x) do { } while (0)
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache
*cachep
)
520 return cachep
->obj_offset
;
523 static int obj_size(struct kmem_cache
*cachep
)
525 return cachep
->obj_size
;
528 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
530 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
531 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
534 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
536 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
537 if (cachep
->flags
& SLAB_STORE_USER
)
538 return (unsigned long *)(objp
+ cachep
->buffer_size
-
540 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
543 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
545 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
546 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
588 page
->lru
.next
= (struct list_head
*)cache
;
591 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
593 if (unlikely(PageCompound(page
)))
594 page
= (struct page
*)page_private(page
);
595 BUG_ON(!PageSlab(page
));
596 return (struct kmem_cache
*)page
->lru
.next
;
599 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
601 page
->lru
.prev
= (struct list_head
*)slab
;
604 static inline struct slab
*page_get_slab(struct page
*page
)
606 if (unlikely(PageCompound(page
)))
607 page
= (struct page
*)page_private(page
);
608 BUG_ON(!PageSlab(page
));
609 return (struct slab
*)page
->lru
.prev
;
612 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
614 struct page
*page
= virt_to_page(obj
);
615 return page_get_cache(page
);
618 static inline struct slab
*virt_to_slab(const void *obj
)
620 struct page
*page
= virt_to_page(obj
);
621 return page_get_slab(page
);
624 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
627 return slab
->s_mem
+ cache
->buffer_size
* idx
;
630 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
631 struct slab
*slab
, void *obj
)
633 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 struct cache_sizes malloc_sizes
[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
645 EXPORT_SYMBOL(malloc_sizes
);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
653 static struct cache_names __initdata cache_names
[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
660 static struct arraycache_init initarray_cache __initdata
=
661 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
662 static struct arraycache_init initarray_generic
=
663 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache
= {
668 .limit
= BOOT_CPUCACHE_ENTRIES
,
670 .buffer_size
= sizeof(struct kmem_cache
),
671 .name
= "kmem_cache",
673 .obj_size
= sizeof(struct kmem_cache
),
677 #ifdef CONFIG_LOCKDEP
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
686 static struct lock_class_key on_slab_key
;
688 static inline void init_lock_keys(struct cache_sizes
*s
)
692 for (q
= 0; q
< MAX_NUMNODES
; q
++) {
693 if (!s
->cs_cachep
->nodelists
[q
] || OFF_SLAB(s
->cs_cachep
))
695 lockdep_set_class(&s
->cs_cachep
->nodelists
[q
]->list_lock
,
701 static inline void init_lock_keys(struct cache_sizes
*s
)
708 /* Guard access to the cache-chain. */
709 static DEFINE_MUTEX(cache_chain_mutex
);
710 static struct list_head cache_chain
;
713 * vm_enough_memory() looks at this to determine how many slab-allocated pages
714 * are possibly freeable under pressure
716 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
718 atomic_t slab_reclaim_pages
;
721 * chicken and egg problem: delay the per-cpu array allocation
722 * until the general caches are up.
732 * used by boot code to determine if it can use slab based allocator
734 int slab_is_available(void)
736 return g_cpucache_up
== FULL
;
739 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
741 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
743 return cachep
->array
[smp_processor_id()];
746 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
749 struct cache_sizes
*csizep
= malloc_sizes
;
752 /* This happens if someone tries to call
753 * kmem_cache_create(), or __kmalloc(), before
754 * the generic caches are initialized.
756 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
758 while (size
> csizep
->cs_size
)
762 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
763 * has cs_{dma,}cachep==NULL. Thus no special case
764 * for large kmalloc calls required.
766 if (unlikely(gfpflags
& GFP_DMA
))
767 return csizep
->cs_dmacachep
;
768 return csizep
->cs_cachep
;
771 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
773 return __find_general_cachep(size
, gfpflags
);
776 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
778 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
782 * Calculate the number of objects and left-over bytes for a given buffer size.
784 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
785 size_t align
, int flags
, size_t *left_over
,
790 size_t slab_size
= PAGE_SIZE
<< gfporder
;
793 * The slab management structure can be either off the slab or
794 * on it. For the latter case, the memory allocated for a
798 * - One kmem_bufctl_t for each object
799 * - Padding to respect alignment of @align
800 * - @buffer_size bytes for each object
802 * If the slab management structure is off the slab, then the
803 * alignment will already be calculated into the size. Because
804 * the slabs are all pages aligned, the objects will be at the
805 * correct alignment when allocated.
807 if (flags
& CFLGS_OFF_SLAB
) {
809 nr_objs
= slab_size
/ buffer_size
;
811 if (nr_objs
> SLAB_LIMIT
)
812 nr_objs
= SLAB_LIMIT
;
815 * Ignore padding for the initial guess. The padding
816 * is at most @align-1 bytes, and @buffer_size is at
817 * least @align. In the worst case, this result will
818 * be one greater than the number of objects that fit
819 * into the memory allocation when taking the padding
822 nr_objs
= (slab_size
- sizeof(struct slab
)) /
823 (buffer_size
+ sizeof(kmem_bufctl_t
));
826 * This calculated number will be either the right
827 * amount, or one greater than what we want.
829 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
833 if (nr_objs
> SLAB_LIMIT
)
834 nr_objs
= SLAB_LIMIT
;
836 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
839 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
842 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
844 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
847 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
848 function
, cachep
->name
, msg
);
854 * Special reaping functions for NUMA systems called from cache_reap().
855 * These take care of doing round robin flushing of alien caches (containing
856 * objects freed on different nodes from which they were allocated) and the
857 * flushing of remote pcps by calling drain_node_pages.
859 static DEFINE_PER_CPU(unsigned long, reap_node
);
861 static void init_reap_node(int cpu
)
865 node
= next_node(cpu_to_node(cpu
), node_online_map
);
866 if (node
== MAX_NUMNODES
)
867 node
= first_node(node_online_map
);
869 __get_cpu_var(reap_node
) = node
;
872 static void next_reap_node(void)
874 int node
= __get_cpu_var(reap_node
);
877 * Also drain per cpu pages on remote zones
879 if (node
!= numa_node_id())
880 drain_node_pages(node
);
882 node
= next_node(node
, node_online_map
);
883 if (unlikely(node
>= MAX_NUMNODES
))
884 node
= first_node(node_online_map
);
885 __get_cpu_var(reap_node
) = node
;
889 #define init_reap_node(cpu) do { } while (0)
890 #define next_reap_node(void) do { } while (0)
894 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
895 * via the workqueue/eventd.
896 * Add the CPU number into the expiration time to minimize the possibility of
897 * the CPUs getting into lockstep and contending for the global cache chain
900 static void __devinit
start_cpu_timer(int cpu
)
902 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
905 * When this gets called from do_initcalls via cpucache_init(),
906 * init_workqueues() has already run, so keventd will be setup
909 if (keventd_up() && reap_work
->func
== NULL
) {
911 INIT_WORK(reap_work
, cache_reap
, NULL
);
912 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
916 static struct array_cache
*alloc_arraycache(int node
, int entries
,
919 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
920 struct array_cache
*nc
= NULL
;
922 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
926 nc
->batchcount
= batchcount
;
928 spin_lock_init(&nc
->lock
);
934 * Transfer objects in one arraycache to another.
935 * Locking must be handled by the caller.
937 * Return the number of entries transferred.
939 static int transfer_objects(struct array_cache
*to
,
940 struct array_cache
*from
, unsigned int max
)
942 /* Figure out how many entries to transfer */
943 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
948 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
958 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
959 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
961 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
963 struct array_cache
**ac_ptr
;
964 int memsize
= sizeof(void *) * MAX_NUMNODES
;
969 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
972 if (i
== node
|| !node_online(i
)) {
976 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
978 for (i
--; i
<= 0; i
--)
988 static void free_alien_cache(struct array_cache
**ac_ptr
)
999 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1000 struct array_cache
*ac
, int node
)
1002 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1005 spin_lock(&rl3
->list_lock
);
1007 * Stuff objects into the remote nodes shared array first.
1008 * That way we could avoid the overhead of putting the objects
1009 * into the free lists and getting them back later.
1012 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1014 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1016 spin_unlock(&rl3
->list_lock
);
1021 * Called from cache_reap() to regularly drain alien caches round robin.
1023 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1025 int node
= __get_cpu_var(reap_node
);
1028 struct array_cache
*ac
= l3
->alien
[node
];
1030 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1031 __drain_alien_cache(cachep
, ac
, node
);
1032 spin_unlock_irq(&ac
->lock
);
1037 static void drain_alien_cache(struct kmem_cache
*cachep
,
1038 struct array_cache
**alien
)
1041 struct array_cache
*ac
;
1042 unsigned long flags
;
1044 for_each_online_node(i
) {
1047 spin_lock_irqsave(&ac
->lock
, flags
);
1048 __drain_alien_cache(cachep
, ac
, i
);
1049 spin_unlock_irqrestore(&ac
->lock
, flags
);
1054 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1056 struct slab
*slabp
= virt_to_slab(objp
);
1057 int nodeid
= slabp
->nodeid
;
1058 struct kmem_list3
*l3
;
1059 struct array_cache
*alien
= NULL
;
1062 * Make sure we are not freeing a object from another node to the array
1063 * cache on this cpu.
1065 if (likely(slabp
->nodeid
== numa_node_id()))
1068 l3
= cachep
->nodelists
[numa_node_id()];
1069 STATS_INC_NODEFREES(cachep
);
1070 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1071 alien
= l3
->alien
[nodeid
];
1072 spin_lock(&alien
->lock
);
1073 if (unlikely(alien
->avail
== alien
->limit
)) {
1074 STATS_INC_ACOVERFLOW(cachep
);
1075 __drain_alien_cache(cachep
, alien
, nodeid
);
1077 alien
->entry
[alien
->avail
++] = objp
;
1078 spin_unlock(&alien
->lock
);
1080 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1081 free_block(cachep
, &objp
, 1, nodeid
);
1082 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1089 #define drain_alien_cache(cachep, alien) do { } while (0)
1090 #define reap_alien(cachep, l3) do { } while (0)
1092 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1094 return (struct array_cache
**) 0x01020304ul
;
1097 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1101 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1108 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1109 unsigned long action
, void *hcpu
)
1111 long cpu
= (long)hcpu
;
1112 struct kmem_cache
*cachep
;
1113 struct kmem_list3
*l3
= NULL
;
1114 int node
= cpu_to_node(cpu
);
1115 int memsize
= sizeof(struct kmem_list3
);
1118 case CPU_UP_PREPARE
:
1119 mutex_lock(&cache_chain_mutex
);
1121 * We need to do this right in the beginning since
1122 * alloc_arraycache's are going to use this list.
1123 * kmalloc_node allows us to add the slab to the right
1124 * kmem_list3 and not this cpu's kmem_list3
1127 list_for_each_entry(cachep
, &cache_chain
, next
) {
1129 * Set up the size64 kmemlist for cpu before we can
1130 * begin anything. Make sure some other cpu on this
1131 * node has not already allocated this
1133 if (!cachep
->nodelists
[node
]) {
1134 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1137 kmem_list3_init(l3
);
1138 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1139 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1142 * The l3s don't come and go as CPUs come and
1143 * go. cache_chain_mutex is sufficient
1146 cachep
->nodelists
[node
] = l3
;
1149 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1150 cachep
->nodelists
[node
]->free_limit
=
1151 (1 + nr_cpus_node(node
)) *
1152 cachep
->batchcount
+ cachep
->num
;
1153 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1157 * Now we can go ahead with allocating the shared arrays and
1160 list_for_each_entry(cachep
, &cache_chain
, next
) {
1161 struct array_cache
*nc
;
1162 struct array_cache
*shared
;
1163 struct array_cache
**alien
;
1165 nc
= alloc_arraycache(node
, cachep
->limit
,
1166 cachep
->batchcount
);
1169 shared
= alloc_arraycache(node
,
1170 cachep
->shared
* cachep
->batchcount
,
1175 alien
= alloc_alien_cache(node
, cachep
->limit
);
1178 cachep
->array
[cpu
] = nc
;
1179 l3
= cachep
->nodelists
[node
];
1182 spin_lock_irq(&l3
->list_lock
);
1185 * We are serialised from CPU_DEAD or
1186 * CPU_UP_CANCELLED by the cpucontrol lock
1188 l3
->shared
= shared
;
1197 spin_unlock_irq(&l3
->list_lock
);
1199 free_alien_cache(alien
);
1201 mutex_unlock(&cache_chain_mutex
);
1204 start_cpu_timer(cpu
);
1206 #ifdef CONFIG_HOTPLUG_CPU
1209 * Even if all the cpus of a node are down, we don't free the
1210 * kmem_list3 of any cache. This to avoid a race between
1211 * cpu_down, and a kmalloc allocation from another cpu for
1212 * memory from the node of the cpu going down. The list3
1213 * structure is usually allocated from kmem_cache_create() and
1214 * gets destroyed at kmem_cache_destroy().
1217 case CPU_UP_CANCELED
:
1218 mutex_lock(&cache_chain_mutex
);
1219 list_for_each_entry(cachep
, &cache_chain
, next
) {
1220 struct array_cache
*nc
;
1221 struct array_cache
*shared
;
1222 struct array_cache
**alien
;
1225 mask
= node_to_cpumask(node
);
1226 /* cpu is dead; no one can alloc from it. */
1227 nc
= cachep
->array
[cpu
];
1228 cachep
->array
[cpu
] = NULL
;
1229 l3
= cachep
->nodelists
[node
];
1232 goto free_array_cache
;
1234 spin_lock_irq(&l3
->list_lock
);
1236 /* Free limit for this kmem_list3 */
1237 l3
->free_limit
-= cachep
->batchcount
;
1239 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1241 if (!cpus_empty(mask
)) {
1242 spin_unlock_irq(&l3
->list_lock
);
1243 goto free_array_cache
;
1246 shared
= l3
->shared
;
1248 free_block(cachep
, l3
->shared
->entry
,
1249 l3
->shared
->avail
, node
);
1256 spin_unlock_irq(&l3
->list_lock
);
1260 drain_alien_cache(cachep
, alien
);
1261 free_alien_cache(alien
);
1267 * In the previous loop, all the objects were freed to
1268 * the respective cache's slabs, now we can go ahead and
1269 * shrink each nodelist to its limit.
1271 list_for_each_entry(cachep
, &cache_chain
, next
) {
1272 l3
= cachep
->nodelists
[node
];
1275 drain_freelist(cachep
, l3
, l3
->free_objects
);
1277 mutex_unlock(&cache_chain_mutex
);
1283 mutex_unlock(&cache_chain_mutex
);
1287 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1288 &cpuup_callback
, NULL
, 0
1292 * swap the static kmem_list3 with kmalloced memory
1294 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1297 struct kmem_list3
*ptr
;
1299 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1300 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1303 local_irq_disable();
1304 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1306 * Do not assume that spinlocks can be initialized via memcpy:
1308 spin_lock_init(&ptr
->list_lock
);
1310 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1311 cachep
->nodelists
[nodeid
] = ptr
;
1316 * Initialisation. Called after the page allocator have been initialised and
1317 * before smp_init().
1319 void __init
kmem_cache_init(void)
1322 struct cache_sizes
*sizes
;
1323 struct cache_names
*names
;
1327 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1328 kmem_list3_init(&initkmem_list3
[i
]);
1329 if (i
< MAX_NUMNODES
)
1330 cache_cache
.nodelists
[i
] = NULL
;
1334 * Fragmentation resistance on low memory - only use bigger
1335 * page orders on machines with more than 32MB of memory.
1337 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1338 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1340 /* Bootstrap is tricky, because several objects are allocated
1341 * from caches that do not exist yet:
1342 * 1) initialize the cache_cache cache: it contains the struct
1343 * kmem_cache structures of all caches, except cache_cache itself:
1344 * cache_cache is statically allocated.
1345 * Initially an __init data area is used for the head array and the
1346 * kmem_list3 structures, it's replaced with a kmalloc allocated
1347 * array at the end of the bootstrap.
1348 * 2) Create the first kmalloc cache.
1349 * The struct kmem_cache for the new cache is allocated normally.
1350 * An __init data area is used for the head array.
1351 * 3) Create the remaining kmalloc caches, with minimally sized
1353 * 4) Replace the __init data head arrays for cache_cache and the first
1354 * kmalloc cache with kmalloc allocated arrays.
1355 * 5) Replace the __init data for kmem_list3 for cache_cache and
1356 * the other cache's with kmalloc allocated memory.
1357 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1360 /* 1) create the cache_cache */
1361 INIT_LIST_HEAD(&cache_chain
);
1362 list_add(&cache_cache
.next
, &cache_chain
);
1363 cache_cache
.colour_off
= cache_line_size();
1364 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1365 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1367 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1370 for (order
= 0; order
< MAX_ORDER
; order
++) {
1371 cache_estimate(order
, cache_cache
.buffer_size
,
1372 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1373 if (cache_cache
.num
)
1376 BUG_ON(!cache_cache
.num
);
1377 cache_cache
.gfporder
= order
;
1378 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1379 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1380 sizeof(struct slab
), cache_line_size());
1382 /* 2+3) create the kmalloc caches */
1383 sizes
= malloc_sizes
;
1384 names
= cache_names
;
1387 * Initialize the caches that provide memory for the array cache and the
1388 * kmem_list3 structures first. Without this, further allocations will
1392 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1393 sizes
[INDEX_AC
].cs_size
,
1394 ARCH_KMALLOC_MINALIGN
,
1395 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1398 if (INDEX_AC
!= INDEX_L3
) {
1399 sizes
[INDEX_L3
].cs_cachep
=
1400 kmem_cache_create(names
[INDEX_L3
].name
,
1401 sizes
[INDEX_L3
].cs_size
,
1402 ARCH_KMALLOC_MINALIGN
,
1403 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1407 slab_early_init
= 0;
1409 while (sizes
->cs_size
!= ULONG_MAX
) {
1411 * For performance, all the general caches are L1 aligned.
1412 * This should be particularly beneficial on SMP boxes, as it
1413 * eliminates "false sharing".
1414 * Note for systems short on memory removing the alignment will
1415 * allow tighter packing of the smaller caches.
1417 if (!sizes
->cs_cachep
) {
1418 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1420 ARCH_KMALLOC_MINALIGN
,
1421 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1424 init_lock_keys(sizes
);
1426 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1428 ARCH_KMALLOC_MINALIGN
,
1429 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1435 /* 4) Replace the bootstrap head arrays */
1437 struct array_cache
*ptr
;
1439 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1441 local_irq_disable();
1442 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1443 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1444 sizeof(struct arraycache_init
));
1446 * Do not assume that spinlocks can be initialized via memcpy:
1448 spin_lock_init(&ptr
->lock
);
1450 cache_cache
.array
[smp_processor_id()] = ptr
;
1453 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1455 local_irq_disable();
1456 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1457 != &initarray_generic
.cache
);
1458 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1459 sizeof(struct arraycache_init
));
1461 * Do not assume that spinlocks can be initialized via memcpy:
1463 spin_lock_init(&ptr
->lock
);
1465 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1469 /* 5) Replace the bootstrap kmem_list3's */
1472 /* Replace the static kmem_list3 structures for the boot cpu */
1473 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1476 for_each_online_node(node
) {
1477 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1478 &initkmem_list3
[SIZE_AC
+ node
], node
);
1480 if (INDEX_AC
!= INDEX_L3
) {
1481 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1482 &initkmem_list3
[SIZE_L3
+ node
],
1488 /* 6) resize the head arrays to their final sizes */
1490 struct kmem_cache
*cachep
;
1491 mutex_lock(&cache_chain_mutex
);
1492 list_for_each_entry(cachep
, &cache_chain
, next
)
1493 enable_cpucache(cachep
);
1494 mutex_unlock(&cache_chain_mutex
);
1498 g_cpucache_up
= FULL
;
1501 * Register a cpu startup notifier callback that initializes
1502 * cpu_cache_get for all new cpus
1504 register_cpu_notifier(&cpucache_notifier
);
1507 * The reap timers are started later, with a module init call: That part
1508 * of the kernel is not yet operational.
1512 static int __init
cpucache_init(void)
1517 * Register the timers that return unneeded pages to the page allocator
1519 for_each_online_cpu(cpu
)
1520 start_cpu_timer(cpu
);
1523 __initcall(cpucache_init
);
1526 * Interface to system's page allocator. No need to hold the cache-lock.
1528 * If we requested dmaable memory, we will get it. Even if we
1529 * did not request dmaable memory, we might get it, but that
1530 * would be relatively rare and ignorable.
1532 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1540 * Nommu uses slab's for process anonymous memory allocations, and thus
1541 * requires __GFP_COMP to properly refcount higher order allocations
1543 flags
|= __GFP_COMP
;
1545 flags
|= cachep
->gfpflags
;
1547 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1551 nr_pages
= (1 << cachep
->gfporder
);
1552 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1553 atomic_add(nr_pages
, &slab_reclaim_pages
);
1554 add_zone_page_state(page_zone(page
), NR_SLAB
, nr_pages
);
1555 for (i
= 0; i
< nr_pages
; i
++)
1556 __SetPageSlab(page
+ i
);
1557 return page_address(page
);
1561 * Interface to system's page release.
1563 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1565 unsigned long i
= (1 << cachep
->gfporder
);
1566 struct page
*page
= virt_to_page(addr
);
1567 const unsigned long nr_freed
= i
;
1569 sub_zone_page_state(page_zone(page
), NR_SLAB
, nr_freed
);
1571 BUG_ON(!PageSlab(page
));
1572 __ClearPageSlab(page
);
1575 if (current
->reclaim_state
)
1576 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1577 free_pages((unsigned long)addr
, cachep
->gfporder
);
1578 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1579 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1582 static void kmem_rcu_free(struct rcu_head
*head
)
1584 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1585 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1587 kmem_freepages(cachep
, slab_rcu
->addr
);
1588 if (OFF_SLAB(cachep
))
1589 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1594 #ifdef CONFIG_DEBUG_PAGEALLOC
1595 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1596 unsigned long caller
)
1598 int size
= obj_size(cachep
);
1600 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1602 if (size
< 5 * sizeof(unsigned long))
1605 *addr
++ = 0x12345678;
1607 *addr
++ = smp_processor_id();
1608 size
-= 3 * sizeof(unsigned long);
1610 unsigned long *sptr
= &caller
;
1611 unsigned long svalue
;
1613 while (!kstack_end(sptr
)) {
1615 if (kernel_text_address(svalue
)) {
1617 size
-= sizeof(unsigned long);
1618 if (size
<= sizeof(unsigned long))
1624 *addr
++ = 0x87654321;
1628 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1630 int size
= obj_size(cachep
);
1631 addr
= &((char *)addr
)[obj_offset(cachep
)];
1633 memset(addr
, val
, size
);
1634 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1637 static void dump_line(char *data
, int offset
, int limit
)
1640 printk(KERN_ERR
"%03x:", offset
);
1641 for (i
= 0; i
< limit
; i
++)
1642 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1649 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1654 if (cachep
->flags
& SLAB_RED_ZONE
) {
1655 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1656 *dbg_redzone1(cachep
, objp
),
1657 *dbg_redzone2(cachep
, objp
));
1660 if (cachep
->flags
& SLAB_STORE_USER
) {
1661 printk(KERN_ERR
"Last user: [<%p>]",
1662 *dbg_userword(cachep
, objp
));
1663 print_symbol("(%s)",
1664 (unsigned long)*dbg_userword(cachep
, objp
));
1667 realobj
= (char *)objp
+ obj_offset(cachep
);
1668 size
= obj_size(cachep
);
1669 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1672 if (i
+ limit
> size
)
1674 dump_line(realobj
, i
, limit
);
1678 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1684 realobj
= (char *)objp
+ obj_offset(cachep
);
1685 size
= obj_size(cachep
);
1687 for (i
= 0; i
< size
; i
++) {
1688 char exp
= POISON_FREE
;
1691 if (realobj
[i
] != exp
) {
1697 "Slab corruption: start=%p, len=%d\n",
1699 print_objinfo(cachep
, objp
, 0);
1701 /* Hexdump the affected line */
1704 if (i
+ limit
> size
)
1706 dump_line(realobj
, i
, limit
);
1709 /* Limit to 5 lines */
1715 /* Print some data about the neighboring objects, if they
1718 struct slab
*slabp
= virt_to_slab(objp
);
1721 objnr
= obj_to_index(cachep
, slabp
, objp
);
1723 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1724 realobj
= (char *)objp
+ obj_offset(cachep
);
1725 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1727 print_objinfo(cachep
, objp
, 2);
1729 if (objnr
+ 1 < cachep
->num
) {
1730 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1731 realobj
= (char *)objp
+ obj_offset(cachep
);
1732 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1734 print_objinfo(cachep
, objp
, 2);
1742 * slab_destroy_objs - destroy a slab and its objects
1743 * @cachep: cache pointer being destroyed
1744 * @slabp: slab pointer being destroyed
1746 * Call the registered destructor for each object in a slab that is being
1749 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1752 for (i
= 0; i
< cachep
->num
; i
++) {
1753 void *objp
= index_to_obj(cachep
, slabp
, i
);
1755 if (cachep
->flags
& SLAB_POISON
) {
1756 #ifdef CONFIG_DEBUG_PAGEALLOC
1757 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1759 kernel_map_pages(virt_to_page(objp
),
1760 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1762 check_poison_obj(cachep
, objp
);
1764 check_poison_obj(cachep
, objp
);
1767 if (cachep
->flags
& SLAB_RED_ZONE
) {
1768 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1769 slab_error(cachep
, "start of a freed object "
1771 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1772 slab_error(cachep
, "end of a freed object "
1775 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1776 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1780 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1784 for (i
= 0; i
< cachep
->num
; i
++) {
1785 void *objp
= index_to_obj(cachep
, slabp
, i
);
1786 (cachep
->dtor
) (objp
, cachep
, 0);
1793 * slab_destroy - destroy and release all objects in a slab
1794 * @cachep: cache pointer being destroyed
1795 * @slabp: slab pointer being destroyed
1797 * Destroy all the objs in a slab, and release the mem back to the system.
1798 * Before calling the slab must have been unlinked from the cache. The
1799 * cache-lock is not held/needed.
1801 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1803 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1805 slab_destroy_objs(cachep
, slabp
);
1806 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1807 struct slab_rcu
*slab_rcu
;
1809 slab_rcu
= (struct slab_rcu
*)slabp
;
1810 slab_rcu
->cachep
= cachep
;
1811 slab_rcu
->addr
= addr
;
1812 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1814 kmem_freepages(cachep
, addr
);
1815 if (OFF_SLAB(cachep
))
1816 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1821 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1822 * size of kmem_list3.
1824 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1828 for_each_online_node(node
) {
1829 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1830 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1832 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1836 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1839 struct kmem_list3
*l3
;
1841 for_each_online_cpu(i
)
1842 kfree(cachep
->array
[i
]);
1844 /* NUMA: free the list3 structures */
1845 for_each_online_node(i
) {
1846 l3
= cachep
->nodelists
[i
];
1849 free_alien_cache(l3
->alien
);
1853 kmem_cache_free(&cache_cache
, cachep
);
1858 * calculate_slab_order - calculate size (page order) of slabs
1859 * @cachep: pointer to the cache that is being created
1860 * @size: size of objects to be created in this cache.
1861 * @align: required alignment for the objects.
1862 * @flags: slab allocation flags
1864 * Also calculates the number of objects per slab.
1866 * This could be made much more intelligent. For now, try to avoid using
1867 * high order pages for slabs. When the gfp() functions are more friendly
1868 * towards high-order requests, this should be changed.
1870 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1871 size_t size
, size_t align
, unsigned long flags
)
1873 unsigned long offslab_limit
;
1874 size_t left_over
= 0;
1877 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1881 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1885 if (flags
& CFLGS_OFF_SLAB
) {
1887 * Max number of objs-per-slab for caches which
1888 * use off-slab slabs. Needed to avoid a possible
1889 * looping condition in cache_grow().
1891 offslab_limit
= size
- sizeof(struct slab
);
1892 offslab_limit
/= sizeof(kmem_bufctl_t
);
1894 if (num
> offslab_limit
)
1898 /* Found something acceptable - save it away */
1900 cachep
->gfporder
= gfporder
;
1901 left_over
= remainder
;
1904 * A VFS-reclaimable slab tends to have most allocations
1905 * as GFP_NOFS and we really don't want to have to be allocating
1906 * higher-order pages when we are unable to shrink dcache.
1908 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1912 * Large number of objects is good, but very large slabs are
1913 * currently bad for the gfp()s.
1915 if (gfporder
>= slab_break_gfp_order
)
1919 * Acceptable internal fragmentation?
1921 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1927 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1929 if (g_cpucache_up
== FULL
) {
1930 enable_cpucache(cachep
);
1933 if (g_cpucache_up
== NONE
) {
1935 * Note: the first kmem_cache_create must create the cache
1936 * that's used by kmalloc(24), otherwise the creation of
1937 * further caches will BUG().
1939 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1942 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1943 * the first cache, then we need to set up all its list3s,
1944 * otherwise the creation of further caches will BUG().
1946 set_up_list3s(cachep
, SIZE_AC
);
1947 if (INDEX_AC
== INDEX_L3
)
1948 g_cpucache_up
= PARTIAL_L3
;
1950 g_cpucache_up
= PARTIAL_AC
;
1952 cachep
->array
[smp_processor_id()] =
1953 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1955 if (g_cpucache_up
== PARTIAL_AC
) {
1956 set_up_list3s(cachep
, SIZE_L3
);
1957 g_cpucache_up
= PARTIAL_L3
;
1960 for_each_online_node(node
) {
1961 cachep
->nodelists
[node
] =
1962 kmalloc_node(sizeof(struct kmem_list3
),
1964 BUG_ON(!cachep
->nodelists
[node
]);
1965 kmem_list3_init(cachep
->nodelists
[node
]);
1969 cachep
->nodelists
[numa_node_id()]->next_reap
=
1970 jiffies
+ REAPTIMEOUT_LIST3
+
1971 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1973 cpu_cache_get(cachep
)->avail
= 0;
1974 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1975 cpu_cache_get(cachep
)->batchcount
= 1;
1976 cpu_cache_get(cachep
)->touched
= 0;
1977 cachep
->batchcount
= 1;
1978 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1982 * kmem_cache_create - Create a cache.
1983 * @name: A string which is used in /proc/slabinfo to identify this cache.
1984 * @size: The size of objects to be created in this cache.
1985 * @align: The required alignment for the objects.
1986 * @flags: SLAB flags
1987 * @ctor: A constructor for the objects.
1988 * @dtor: A destructor for the objects.
1990 * Returns a ptr to the cache on success, NULL on failure.
1991 * Cannot be called within a int, but can be interrupted.
1992 * The @ctor is run when new pages are allocated by the cache
1993 * and the @dtor is run before the pages are handed back.
1995 * @name must be valid until the cache is destroyed. This implies that
1996 * the module calling this has to destroy the cache before getting unloaded.
2000 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2001 * to catch references to uninitialised memory.
2003 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2004 * for buffer overruns.
2006 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2007 * cacheline. This can be beneficial if you're counting cycles as closely
2011 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2012 unsigned long flags
,
2013 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
2014 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
2016 size_t left_over
, slab_size
, ralign
;
2017 struct kmem_cache
*cachep
= NULL
, *pc
;
2020 * Sanity checks... these are all serious usage bugs.
2022 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2023 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
2024 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
2030 * Prevent CPUs from coming and going.
2031 * lock_cpu_hotplug() nests outside cache_chain_mutex
2035 mutex_lock(&cache_chain_mutex
);
2037 list_for_each_entry(pc
, &cache_chain
, next
) {
2038 mm_segment_t old_fs
= get_fs();
2043 * This happens when the module gets unloaded and doesn't
2044 * destroy its slab cache and no-one else reuses the vmalloc
2045 * area of the module. Print a warning.
2048 res
= __get_user(tmp
, pc
->name
);
2051 printk("SLAB: cache with size %d has lost its name\n",
2056 if (!strcmp(pc
->name
, name
)) {
2057 printk("kmem_cache_create: duplicate cache %s\n", name
);
2064 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2065 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
2066 /* No constructor, but inital state check requested */
2067 printk(KERN_ERR
"%s: No con, but init state check "
2068 "requested - %s\n", __FUNCTION__
, name
);
2069 flags
&= ~SLAB_DEBUG_INITIAL
;
2073 * Enable redzoning and last user accounting, except for caches with
2074 * large objects, if the increased size would increase the object size
2075 * above the next power of two: caches with object sizes just above a
2076 * power of two have a significant amount of internal fragmentation.
2078 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2079 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2080 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2081 flags
|= SLAB_POISON
;
2083 if (flags
& SLAB_DESTROY_BY_RCU
)
2084 BUG_ON(flags
& SLAB_POISON
);
2086 if (flags
& SLAB_DESTROY_BY_RCU
)
2090 * Always checks flags, a caller might be expecting debug support which
2093 BUG_ON(flags
& ~CREATE_MASK
);
2096 * Check that size is in terms of words. This is needed to avoid
2097 * unaligned accesses for some archs when redzoning is used, and makes
2098 * sure any on-slab bufctl's are also correctly aligned.
2100 if (size
& (BYTES_PER_WORD
- 1)) {
2101 size
+= (BYTES_PER_WORD
- 1);
2102 size
&= ~(BYTES_PER_WORD
- 1);
2105 /* calculate the final buffer alignment: */
2107 /* 1) arch recommendation: can be overridden for debug */
2108 if (flags
& SLAB_HWCACHE_ALIGN
) {
2110 * Default alignment: as specified by the arch code. Except if
2111 * an object is really small, then squeeze multiple objects into
2114 ralign
= cache_line_size();
2115 while (size
<= ralign
/ 2)
2118 ralign
= BYTES_PER_WORD
;
2122 * Redzoning and user store require word alignment. Note this will be
2123 * overridden by architecture or caller mandated alignment if either
2124 * is greater than BYTES_PER_WORD.
2126 if (flags
& SLAB_RED_ZONE
|| flags
& SLAB_STORE_USER
)
2127 ralign
= BYTES_PER_WORD
;
2129 /* 2) arch mandated alignment: disables debug if necessary */
2130 if (ralign
< ARCH_SLAB_MINALIGN
) {
2131 ralign
= ARCH_SLAB_MINALIGN
;
2132 if (ralign
> BYTES_PER_WORD
)
2133 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2135 /* 3) caller mandated alignment: disables debug if necessary */
2136 if (ralign
< align
) {
2138 if (ralign
> BYTES_PER_WORD
)
2139 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2146 /* Get cache's description obj. */
2147 cachep
= kmem_cache_zalloc(&cache_cache
, SLAB_KERNEL
);
2152 cachep
->obj_size
= size
;
2155 * Both debugging options require word-alignment which is calculated
2158 if (flags
& SLAB_RED_ZONE
) {
2159 /* add space for red zone words */
2160 cachep
->obj_offset
+= BYTES_PER_WORD
;
2161 size
+= 2 * BYTES_PER_WORD
;
2163 if (flags
& SLAB_STORE_USER
) {
2164 /* user store requires one word storage behind the end of
2167 size
+= BYTES_PER_WORD
;
2169 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2170 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2171 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2172 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2179 * Determine if the slab management is 'on' or 'off' slab.
2180 * (bootstrapping cannot cope with offslab caches so don't do
2183 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2185 * Size is large, assume best to place the slab management obj
2186 * off-slab (should allow better packing of objs).
2188 flags
|= CFLGS_OFF_SLAB
;
2190 size
= ALIGN(size
, align
);
2192 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2195 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2196 kmem_cache_free(&cache_cache
, cachep
);
2200 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2201 + sizeof(struct slab
), align
);
2204 * If the slab has been placed off-slab, and we have enough space then
2205 * move it on-slab. This is at the expense of any extra colouring.
2207 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2208 flags
&= ~CFLGS_OFF_SLAB
;
2209 left_over
-= slab_size
;
2212 if (flags
& CFLGS_OFF_SLAB
) {
2213 /* really off slab. No need for manual alignment */
2215 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2218 cachep
->colour_off
= cache_line_size();
2219 /* Offset must be a multiple of the alignment. */
2220 if (cachep
->colour_off
< align
)
2221 cachep
->colour_off
= align
;
2222 cachep
->colour
= left_over
/ cachep
->colour_off
;
2223 cachep
->slab_size
= slab_size
;
2224 cachep
->flags
= flags
;
2225 cachep
->gfpflags
= 0;
2226 if (flags
& SLAB_CACHE_DMA
)
2227 cachep
->gfpflags
|= GFP_DMA
;
2228 cachep
->buffer_size
= size
;
2230 if (flags
& CFLGS_OFF_SLAB
) {
2231 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2233 * This is a possibility for one of the malloc_sizes caches.
2234 * But since we go off slab only for object size greater than
2235 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2236 * this should not happen at all.
2237 * But leave a BUG_ON for some lucky dude.
2239 BUG_ON(!cachep
->slabp_cache
);
2241 cachep
->ctor
= ctor
;
2242 cachep
->dtor
= dtor
;
2243 cachep
->name
= name
;
2246 setup_cpu_cache(cachep
);
2248 /* cache setup completed, link it into the list */
2249 list_add(&cachep
->next
, &cache_chain
);
2251 if (!cachep
&& (flags
& SLAB_PANIC
))
2252 panic("kmem_cache_create(): failed to create slab `%s'\n",
2254 mutex_unlock(&cache_chain_mutex
);
2255 unlock_cpu_hotplug();
2258 EXPORT_SYMBOL(kmem_cache_create
);
2261 static void check_irq_off(void)
2263 BUG_ON(!irqs_disabled());
2266 static void check_irq_on(void)
2268 BUG_ON(irqs_disabled());
2271 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2275 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2279 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2283 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2288 #define check_irq_off() do { } while(0)
2289 #define check_irq_on() do { } while(0)
2290 #define check_spinlock_acquired(x) do { } while(0)
2291 #define check_spinlock_acquired_node(x, y) do { } while(0)
2294 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2295 struct array_cache
*ac
,
2296 int force
, int node
);
2298 static void do_drain(void *arg
)
2300 struct kmem_cache
*cachep
= arg
;
2301 struct array_cache
*ac
;
2302 int node
= numa_node_id();
2305 ac
= cpu_cache_get(cachep
);
2306 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2307 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2308 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2312 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2314 struct kmem_list3
*l3
;
2317 on_each_cpu(do_drain
, cachep
, 1, 1);
2319 for_each_online_node(node
) {
2320 l3
= cachep
->nodelists
[node
];
2321 if (l3
&& l3
->alien
)
2322 drain_alien_cache(cachep
, l3
->alien
);
2325 for_each_online_node(node
) {
2326 l3
= cachep
->nodelists
[node
];
2328 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2333 * Remove slabs from the list of free slabs.
2334 * Specify the number of slabs to drain in tofree.
2336 * Returns the actual number of slabs released.
2338 static int drain_freelist(struct kmem_cache
*cache
,
2339 struct kmem_list3
*l3
, int tofree
)
2341 struct list_head
*p
;
2346 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2348 spin_lock_irq(&l3
->list_lock
);
2349 p
= l3
->slabs_free
.prev
;
2350 if (p
== &l3
->slabs_free
) {
2351 spin_unlock_irq(&l3
->list_lock
);
2355 slabp
= list_entry(p
, struct slab
, list
);
2357 BUG_ON(slabp
->inuse
);
2359 list_del(&slabp
->list
);
2361 * Safe to drop the lock. The slab is no longer linked
2364 l3
->free_objects
-= cache
->num
;
2365 spin_unlock_irq(&l3
->list_lock
);
2366 slab_destroy(cache
, slabp
);
2373 static int __cache_shrink(struct kmem_cache
*cachep
)
2376 struct kmem_list3
*l3
;
2378 drain_cpu_caches(cachep
);
2381 for_each_online_node(i
) {
2382 l3
= cachep
->nodelists
[i
];
2386 drain_freelist(cachep
, l3
, l3
->free_objects
);
2388 ret
+= !list_empty(&l3
->slabs_full
) ||
2389 !list_empty(&l3
->slabs_partial
);
2391 return (ret
? 1 : 0);
2395 * kmem_cache_shrink - Shrink a cache.
2396 * @cachep: The cache to shrink.
2398 * Releases as many slabs as possible for a cache.
2399 * To help debugging, a zero exit status indicates all slabs were released.
2401 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2403 BUG_ON(!cachep
|| in_interrupt());
2405 return __cache_shrink(cachep
);
2407 EXPORT_SYMBOL(kmem_cache_shrink
);
2410 * kmem_cache_destroy - delete a cache
2411 * @cachep: the cache to destroy
2413 * Remove a struct kmem_cache object from the slab cache.
2414 * Returns 0 on success.
2416 * It is expected this function will be called by a module when it is
2417 * unloaded. This will remove the cache completely, and avoid a duplicate
2418 * cache being allocated each time a module is loaded and unloaded, if the
2419 * module doesn't have persistent in-kernel storage across loads and unloads.
2421 * The cache must be empty before calling this function.
2423 * The caller must guarantee that noone will allocate memory from the cache
2424 * during the kmem_cache_destroy().
2426 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2428 BUG_ON(!cachep
|| in_interrupt());
2430 /* Don't let CPUs to come and go */
2433 /* Find the cache in the chain of caches. */
2434 mutex_lock(&cache_chain_mutex
);
2436 * the chain is never empty, cache_cache is never destroyed
2438 list_del(&cachep
->next
);
2439 mutex_unlock(&cache_chain_mutex
);
2441 if (__cache_shrink(cachep
)) {
2442 slab_error(cachep
, "Can't free all objects");
2443 mutex_lock(&cache_chain_mutex
);
2444 list_add(&cachep
->next
, &cache_chain
);
2445 mutex_unlock(&cache_chain_mutex
);
2446 unlock_cpu_hotplug();
2450 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2453 __kmem_cache_destroy(cachep
);
2454 unlock_cpu_hotplug();
2457 EXPORT_SYMBOL(kmem_cache_destroy
);
2460 * Get the memory for a slab management obj.
2461 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2462 * always come from malloc_sizes caches. The slab descriptor cannot
2463 * come from the same cache which is getting created because,
2464 * when we are searching for an appropriate cache for these
2465 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2466 * If we are creating a malloc_sizes cache here it would not be visible to
2467 * kmem_find_general_cachep till the initialization is complete.
2468 * Hence we cannot have slabp_cache same as the original cache.
2470 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2471 int colour_off
, gfp_t local_flags
,
2476 if (OFF_SLAB(cachep
)) {
2477 /* Slab management obj is off-slab. */
2478 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2479 local_flags
, nodeid
);
2483 slabp
= objp
+ colour_off
;
2484 colour_off
+= cachep
->slab_size
;
2487 slabp
->colouroff
= colour_off
;
2488 slabp
->s_mem
= objp
+ colour_off
;
2489 slabp
->nodeid
= nodeid
;
2493 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2495 return (kmem_bufctl_t
*) (slabp
+ 1);
2498 static void cache_init_objs(struct kmem_cache
*cachep
,
2499 struct slab
*slabp
, unsigned long ctor_flags
)
2503 for (i
= 0; i
< cachep
->num
; i
++) {
2504 void *objp
= index_to_obj(cachep
, slabp
, i
);
2506 /* need to poison the objs? */
2507 if (cachep
->flags
& SLAB_POISON
)
2508 poison_obj(cachep
, objp
, POISON_FREE
);
2509 if (cachep
->flags
& SLAB_STORE_USER
)
2510 *dbg_userword(cachep
, objp
) = NULL
;
2512 if (cachep
->flags
& SLAB_RED_ZONE
) {
2513 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2514 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2517 * Constructors are not allowed to allocate memory from the same
2518 * cache which they are a constructor for. Otherwise, deadlock.
2519 * They must also be threaded.
2521 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2522 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2525 if (cachep
->flags
& SLAB_RED_ZONE
) {
2526 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2527 slab_error(cachep
, "constructor overwrote the"
2528 " end of an object");
2529 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2530 slab_error(cachep
, "constructor overwrote the"
2531 " start of an object");
2533 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2534 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2535 kernel_map_pages(virt_to_page(objp
),
2536 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2539 cachep
->ctor(objp
, cachep
, ctor_flags
);
2541 slab_bufctl(slabp
)[i
] = i
+ 1;
2543 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2547 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2549 if (flags
& SLAB_DMA
)
2550 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2552 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2555 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2558 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2562 next
= slab_bufctl(slabp
)[slabp
->free
];
2564 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2565 WARN_ON(slabp
->nodeid
!= nodeid
);
2572 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2573 void *objp
, int nodeid
)
2575 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2578 /* Verify that the slab belongs to the intended node */
2579 WARN_ON(slabp
->nodeid
!= nodeid
);
2581 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2582 printk(KERN_ERR
"slab: double free detected in cache "
2583 "'%s', objp %p\n", cachep
->name
, objp
);
2587 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2588 slabp
->free
= objnr
;
2593 * Map pages beginning at addr to the given cache and slab. This is required
2594 * for the slab allocator to be able to lookup the cache and slab of a
2595 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2597 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2603 page
= virt_to_page(addr
);
2606 if (likely(!PageCompound(page
)))
2607 nr_pages
<<= cache
->gfporder
;
2610 page_set_cache(page
, cache
);
2611 page_set_slab(page
, slab
);
2613 } while (--nr_pages
);
2617 * Grow (by 1) the number of slabs within a cache. This is called by
2618 * kmem_cache_alloc() when there are no active objs left in a cache.
2620 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2626 unsigned long ctor_flags
;
2627 struct kmem_list3
*l3
;
2630 * Be lazy and only check for valid flags here, keeping it out of the
2631 * critical path in kmem_cache_alloc().
2633 BUG_ON(flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
));
2634 if (flags
& SLAB_NO_GROW
)
2637 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2638 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2639 if (!(local_flags
& __GFP_WAIT
))
2641 * Not allowed to sleep. Need to tell a constructor about
2642 * this - it might need to know...
2644 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2646 /* Take the l3 list lock to change the colour_next on this node */
2648 l3
= cachep
->nodelists
[nodeid
];
2649 spin_lock(&l3
->list_lock
);
2651 /* Get colour for the slab, and cal the next value. */
2652 offset
= l3
->colour_next
;
2654 if (l3
->colour_next
>= cachep
->colour
)
2655 l3
->colour_next
= 0;
2656 spin_unlock(&l3
->list_lock
);
2658 offset
*= cachep
->colour_off
;
2660 if (local_flags
& __GFP_WAIT
)
2664 * The test for missing atomic flag is performed here, rather than
2665 * the more obvious place, simply to reduce the critical path length
2666 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2667 * will eventually be caught here (where it matters).
2669 kmem_flagcheck(cachep
, flags
);
2672 * Get mem for the objs. Attempt to allocate a physical page from
2675 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2679 /* Get slab management. */
2680 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
, nodeid
);
2684 slabp
->nodeid
= nodeid
;
2685 slab_map_pages(cachep
, slabp
, objp
);
2687 cache_init_objs(cachep
, slabp
, ctor_flags
);
2689 if (local_flags
& __GFP_WAIT
)
2690 local_irq_disable();
2692 spin_lock(&l3
->list_lock
);
2694 /* Make slab active. */
2695 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2696 STATS_INC_GROWN(cachep
);
2697 l3
->free_objects
+= cachep
->num
;
2698 spin_unlock(&l3
->list_lock
);
2701 kmem_freepages(cachep
, objp
);
2703 if (local_flags
& __GFP_WAIT
)
2704 local_irq_disable();
2711 * Perform extra freeing checks:
2712 * - detect bad pointers.
2713 * - POISON/RED_ZONE checking
2714 * - destructor calls, for caches with POISON+dtor
2716 static void kfree_debugcheck(const void *objp
)
2720 if (!virt_addr_valid(objp
)) {
2721 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2722 (unsigned long)objp
);
2725 page
= virt_to_page(objp
);
2726 if (!PageSlab(page
)) {
2727 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2728 (unsigned long)objp
);
2733 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2735 unsigned long redzone1
, redzone2
;
2737 redzone1
= *dbg_redzone1(cache
, obj
);
2738 redzone2
= *dbg_redzone2(cache
, obj
);
2743 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2746 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2747 slab_error(cache
, "double free detected");
2749 slab_error(cache
, "memory outside object was overwritten");
2751 printk(KERN_ERR
"%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2752 obj
, redzone1
, redzone2
);
2755 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2762 objp
-= obj_offset(cachep
);
2763 kfree_debugcheck(objp
);
2764 page
= virt_to_page(objp
);
2766 slabp
= page_get_slab(page
);
2768 if (cachep
->flags
& SLAB_RED_ZONE
) {
2769 verify_redzone_free(cachep
, objp
);
2770 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2771 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2773 if (cachep
->flags
& SLAB_STORE_USER
)
2774 *dbg_userword(cachep
, objp
) = caller
;
2776 objnr
= obj_to_index(cachep
, slabp
, objp
);
2778 BUG_ON(objnr
>= cachep
->num
);
2779 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2781 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2783 * Need to call the slab's constructor so the caller can
2784 * perform a verify of its state (debugging). Called without
2785 * the cache-lock held.
2787 cachep
->ctor(objp
+ obj_offset(cachep
),
2788 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2790 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2791 /* we want to cache poison the object,
2792 * call the destruction callback
2794 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2796 #ifdef CONFIG_DEBUG_SLAB_LEAK
2797 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2799 if (cachep
->flags
& SLAB_POISON
) {
2800 #ifdef CONFIG_DEBUG_PAGEALLOC
2801 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2802 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2803 kernel_map_pages(virt_to_page(objp
),
2804 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2806 poison_obj(cachep
, objp
, POISON_FREE
);
2809 poison_obj(cachep
, objp
, POISON_FREE
);
2815 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2820 /* Check slab's freelist to see if this obj is there. */
2821 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2823 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2826 if (entries
!= cachep
->num
- slabp
->inuse
) {
2828 printk(KERN_ERR
"slab: Internal list corruption detected in "
2829 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2830 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2832 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2835 printk("\n%03x:", i
);
2836 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2843 #define kfree_debugcheck(x) do { } while(0)
2844 #define cache_free_debugcheck(x,objp,z) (objp)
2845 #define check_slabp(x,y) do { } while(0)
2848 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2851 struct kmem_list3
*l3
;
2852 struct array_cache
*ac
;
2855 ac
= cpu_cache_get(cachep
);
2857 batchcount
= ac
->batchcount
;
2858 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2860 * If there was little recent activity on this cache, then
2861 * perform only a partial refill. Otherwise we could generate
2864 batchcount
= BATCHREFILL_LIMIT
;
2866 l3
= cachep
->nodelists
[numa_node_id()];
2868 BUG_ON(ac
->avail
> 0 || !l3
);
2869 spin_lock(&l3
->list_lock
);
2871 /* See if we can refill from the shared array */
2872 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2875 while (batchcount
> 0) {
2876 struct list_head
*entry
;
2878 /* Get slab alloc is to come from. */
2879 entry
= l3
->slabs_partial
.next
;
2880 if (entry
== &l3
->slabs_partial
) {
2881 l3
->free_touched
= 1;
2882 entry
= l3
->slabs_free
.next
;
2883 if (entry
== &l3
->slabs_free
)
2887 slabp
= list_entry(entry
, struct slab
, list
);
2888 check_slabp(cachep
, slabp
);
2889 check_spinlock_acquired(cachep
);
2890 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2891 STATS_INC_ALLOCED(cachep
);
2892 STATS_INC_ACTIVE(cachep
);
2893 STATS_SET_HIGH(cachep
);
2895 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2898 check_slabp(cachep
, slabp
);
2900 /* move slabp to correct slabp list: */
2901 list_del(&slabp
->list
);
2902 if (slabp
->free
== BUFCTL_END
)
2903 list_add(&slabp
->list
, &l3
->slabs_full
);
2905 list_add(&slabp
->list
, &l3
->slabs_partial
);
2909 l3
->free_objects
-= ac
->avail
;
2911 spin_unlock(&l3
->list_lock
);
2913 if (unlikely(!ac
->avail
)) {
2915 x
= cache_grow(cachep
, flags
, numa_node_id());
2917 /* cache_grow can reenable interrupts, then ac could change. */
2918 ac
= cpu_cache_get(cachep
);
2919 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2922 if (!ac
->avail
) /* objects refilled by interrupt? */
2926 return ac
->entry
[--ac
->avail
];
2929 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2932 might_sleep_if(flags
& __GFP_WAIT
);
2934 kmem_flagcheck(cachep
, flags
);
2939 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2940 gfp_t flags
, void *objp
, void *caller
)
2944 if (cachep
->flags
& SLAB_POISON
) {
2945 #ifdef CONFIG_DEBUG_PAGEALLOC
2946 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2947 kernel_map_pages(virt_to_page(objp
),
2948 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2950 check_poison_obj(cachep
, objp
);
2952 check_poison_obj(cachep
, objp
);
2954 poison_obj(cachep
, objp
, POISON_INUSE
);
2956 if (cachep
->flags
& SLAB_STORE_USER
)
2957 *dbg_userword(cachep
, objp
) = caller
;
2959 if (cachep
->flags
& SLAB_RED_ZONE
) {
2960 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2961 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2962 slab_error(cachep
, "double free, or memory outside"
2963 " object was overwritten");
2965 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2966 objp
, *dbg_redzone1(cachep
, objp
),
2967 *dbg_redzone2(cachep
, objp
));
2969 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2970 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2972 #ifdef CONFIG_DEBUG_SLAB_LEAK
2977 slabp
= page_get_slab(virt_to_page(objp
));
2978 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2979 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
2982 objp
+= obj_offset(cachep
);
2983 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2984 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2986 if (!(flags
& __GFP_WAIT
))
2987 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2989 cachep
->ctor(objp
, cachep
, ctor_flags
);
2994 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2997 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3000 struct array_cache
*ac
;
3003 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3004 objp
= alternate_node_alloc(cachep
, flags
);
3011 ac
= cpu_cache_get(cachep
);
3012 if (likely(ac
->avail
)) {
3013 STATS_INC_ALLOCHIT(cachep
);
3015 objp
= ac
->entry
[--ac
->avail
];
3017 STATS_INC_ALLOCMISS(cachep
);
3018 objp
= cache_alloc_refill(cachep
, flags
);
3023 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
3024 gfp_t flags
, void *caller
)
3026 unsigned long save_flags
;
3029 cache_alloc_debugcheck_before(cachep
, flags
);
3031 local_irq_save(save_flags
);
3032 objp
= ____cache_alloc(cachep
, flags
);
3033 local_irq_restore(save_flags
);
3034 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
3042 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3044 * If we are in_interrupt, then process context, including cpusets and
3045 * mempolicy, may not apply and should not be used for allocation policy.
3047 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3049 int nid_alloc
, nid_here
;
3053 nid_alloc
= nid_here
= numa_node_id();
3054 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3055 nid_alloc
= cpuset_mem_spread_node();
3056 else if (current
->mempolicy
)
3057 nid_alloc
= slab_node(current
->mempolicy
);
3058 if (nid_alloc
!= nid_here
)
3059 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
3064 * A interface to enable slab creation on nodeid
3066 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3069 struct list_head
*entry
;
3071 struct kmem_list3
*l3
;
3075 l3
= cachep
->nodelists
[nodeid
];
3080 spin_lock(&l3
->list_lock
);
3081 entry
= l3
->slabs_partial
.next
;
3082 if (entry
== &l3
->slabs_partial
) {
3083 l3
->free_touched
= 1;
3084 entry
= l3
->slabs_free
.next
;
3085 if (entry
== &l3
->slabs_free
)
3089 slabp
= list_entry(entry
, struct slab
, list
);
3090 check_spinlock_acquired_node(cachep
, nodeid
);
3091 check_slabp(cachep
, slabp
);
3093 STATS_INC_NODEALLOCS(cachep
);
3094 STATS_INC_ACTIVE(cachep
);
3095 STATS_SET_HIGH(cachep
);
3097 BUG_ON(slabp
->inuse
== cachep
->num
);
3099 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3100 check_slabp(cachep
, slabp
);
3102 /* move slabp to correct slabp list: */
3103 list_del(&slabp
->list
);
3105 if (slabp
->free
== BUFCTL_END
)
3106 list_add(&slabp
->list
, &l3
->slabs_full
);
3108 list_add(&slabp
->list
, &l3
->slabs_partial
);
3110 spin_unlock(&l3
->list_lock
);
3114 spin_unlock(&l3
->list_lock
);
3115 x
= cache_grow(cachep
, flags
, nodeid
);
3127 * Caller needs to acquire correct kmem_list's list_lock
3129 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3133 struct kmem_list3
*l3
;
3135 for (i
= 0; i
< nr_objects
; i
++) {
3136 void *objp
= objpp
[i
];
3139 slabp
= virt_to_slab(objp
);
3140 l3
= cachep
->nodelists
[node
];
3141 list_del(&slabp
->list
);
3142 check_spinlock_acquired_node(cachep
, node
);
3143 check_slabp(cachep
, slabp
);
3144 slab_put_obj(cachep
, slabp
, objp
, node
);
3145 STATS_DEC_ACTIVE(cachep
);
3147 check_slabp(cachep
, slabp
);
3149 /* fixup slab chains */
3150 if (slabp
->inuse
== 0) {
3151 if (l3
->free_objects
> l3
->free_limit
) {
3152 l3
->free_objects
-= cachep
->num
;
3153 /* No need to drop any previously held
3154 * lock here, even if we have a off-slab slab
3155 * descriptor it is guaranteed to come from
3156 * a different cache, refer to comments before
3159 slab_destroy(cachep
, slabp
);
3161 list_add(&slabp
->list
, &l3
->slabs_free
);
3164 /* Unconditionally move a slab to the end of the
3165 * partial list on free - maximum time for the
3166 * other objects to be freed, too.
3168 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3173 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3176 struct kmem_list3
*l3
;
3177 int node
= numa_node_id();
3179 batchcount
= ac
->batchcount
;
3181 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3184 l3
= cachep
->nodelists
[node
];
3185 spin_lock(&l3
->list_lock
);
3187 struct array_cache
*shared_array
= l3
->shared
;
3188 int max
= shared_array
->limit
- shared_array
->avail
;
3190 if (batchcount
> max
)
3192 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3193 ac
->entry
, sizeof(void *) * batchcount
);
3194 shared_array
->avail
+= batchcount
;
3199 free_block(cachep
, ac
->entry
, batchcount
, node
);
3204 struct list_head
*p
;
3206 p
= l3
->slabs_free
.next
;
3207 while (p
!= &(l3
->slabs_free
)) {
3210 slabp
= list_entry(p
, struct slab
, list
);
3211 BUG_ON(slabp
->inuse
);
3216 STATS_SET_FREEABLE(cachep
, i
);
3219 spin_unlock(&l3
->list_lock
);
3220 ac
->avail
-= batchcount
;
3221 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3225 * Release an obj back to its cache. If the obj has a constructed state, it must
3226 * be in this state _before_ it is released. Called with disabled ints.
3228 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3230 struct array_cache
*ac
= cpu_cache_get(cachep
);
3233 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3235 if (cache_free_alien(cachep
, objp
))
3238 if (likely(ac
->avail
< ac
->limit
)) {
3239 STATS_INC_FREEHIT(cachep
);
3240 ac
->entry
[ac
->avail
++] = objp
;
3243 STATS_INC_FREEMISS(cachep
);
3244 cache_flusharray(cachep
, ac
);
3245 ac
->entry
[ac
->avail
++] = objp
;
3250 * kmem_cache_alloc - Allocate an object
3251 * @cachep: The cache to allocate from.
3252 * @flags: See kmalloc().
3254 * Allocate an object from this cache. The flags are only relevant
3255 * if the cache has no available objects.
3257 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3259 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3261 EXPORT_SYMBOL(kmem_cache_alloc
);
3264 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3265 * @cache: The cache to allocate from.
3266 * @flags: See kmalloc().
3268 * Allocate an object from this cache and set the allocated memory to zero.
3269 * The flags are only relevant if the cache has no available objects.
3271 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3273 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3275 memset(ret
, 0, obj_size(cache
));
3278 EXPORT_SYMBOL(kmem_cache_zalloc
);
3281 * kmem_ptr_validate - check if an untrusted pointer might
3283 * @cachep: the cache we're checking against
3284 * @ptr: pointer to validate
3286 * This verifies that the untrusted pointer looks sane:
3287 * it is _not_ a guarantee that the pointer is actually
3288 * part of the slab cache in question, but it at least
3289 * validates that the pointer can be dereferenced and
3290 * looks half-way sane.
3292 * Currently only used for dentry validation.
3294 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3296 unsigned long addr
= (unsigned long)ptr
;
3297 unsigned long min_addr
= PAGE_OFFSET
;
3298 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3299 unsigned long size
= cachep
->buffer_size
;
3302 if (unlikely(addr
< min_addr
))
3304 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3306 if (unlikely(addr
& align_mask
))
3308 if (unlikely(!kern_addr_valid(addr
)))
3310 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3312 page
= virt_to_page(ptr
);
3313 if (unlikely(!PageSlab(page
)))
3315 if (unlikely(page_get_cache(page
) != cachep
))
3324 * kmem_cache_alloc_node - Allocate an object on the specified node
3325 * @cachep: The cache to allocate from.
3326 * @flags: See kmalloc().
3327 * @nodeid: node number of the target node.
3329 * Identical to kmem_cache_alloc, except that this function is slow
3330 * and can sleep. And it will allocate memory on the given node, which
3331 * can improve the performance for cpu bound structures.
3332 * New and improved: it will now make sure that the object gets
3333 * put on the correct node list so that there is no false sharing.
3335 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3337 unsigned long save_flags
;
3340 cache_alloc_debugcheck_before(cachep
, flags
);
3341 local_irq_save(save_flags
);
3343 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3344 !cachep
->nodelists
[nodeid
])
3345 ptr
= ____cache_alloc(cachep
, flags
);
3347 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3348 local_irq_restore(save_flags
);
3350 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3351 __builtin_return_address(0));
3355 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3357 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3359 struct kmem_cache
*cachep
;
3361 cachep
= kmem_find_general_cachep(size
, flags
);
3362 if (unlikely(cachep
== NULL
))
3364 return kmem_cache_alloc_node(cachep
, flags
, node
);
3366 EXPORT_SYMBOL(__kmalloc_node
);
3370 * __do_kmalloc - allocate memory
3371 * @size: how many bytes of memory are required.
3372 * @flags: the type of memory to allocate (see kmalloc).
3373 * @caller: function caller for debug tracking of the caller
3375 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3378 struct kmem_cache
*cachep
;
3380 /* If you want to save a few bytes .text space: replace
3382 * Then kmalloc uses the uninlined functions instead of the inline
3385 cachep
= __find_general_cachep(size
, flags
);
3386 if (unlikely(cachep
== NULL
))
3388 return __cache_alloc(cachep
, flags
, caller
);
3392 void *__kmalloc(size_t size
, gfp_t flags
)
3394 #ifndef CONFIG_DEBUG_SLAB
3395 return __do_kmalloc(size
, flags
, NULL
);
3397 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3400 EXPORT_SYMBOL(__kmalloc
);
3402 #ifdef CONFIG_DEBUG_SLAB
3403 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3405 return __do_kmalloc(size
, flags
, caller
);
3407 EXPORT_SYMBOL(__kmalloc_track_caller
);
3412 * percpu_depopulate - depopulate per-cpu data for given cpu
3413 * @__pdata: per-cpu data to depopulate
3414 * @cpu: depopulate per-cpu data for this cpu
3416 * Depopulating per-cpu data for a cpu going offline would be a typical
3417 * use case. You need to register a cpu hotplug handler for that purpose.
3419 void percpu_depopulate(void *__pdata
, int cpu
)
3421 struct percpu_data
*pdata
= __percpu_disguise(__pdata
);
3422 if (pdata
->ptrs
[cpu
]) {
3423 kfree(pdata
->ptrs
[cpu
]);
3424 pdata
->ptrs
[cpu
] = NULL
;
3427 EXPORT_SYMBOL_GPL(percpu_depopulate
);
3430 * percpu_depopulate_mask - depopulate per-cpu data for some cpu's
3431 * @__pdata: per-cpu data to depopulate
3432 * @mask: depopulate per-cpu data for cpu's selected through mask bits
3434 void __percpu_depopulate_mask(void *__pdata
, cpumask_t
*mask
)
3437 for_each_cpu_mask(cpu
, *mask
)
3438 percpu_depopulate(__pdata
, cpu
);
3440 EXPORT_SYMBOL_GPL(__percpu_depopulate_mask
);
3443 * percpu_populate - populate per-cpu data for given cpu
3444 * @__pdata: per-cpu data to populate further
3445 * @size: size of per-cpu object
3446 * @gfp: may sleep or not etc.
3447 * @cpu: populate per-data for this cpu
3449 * Populating per-cpu data for a cpu coming online would be a typical
3450 * use case. You need to register a cpu hotplug handler for that purpose.
3451 * Per-cpu object is populated with zeroed buffer.
3453 void *percpu_populate(void *__pdata
, size_t size
, gfp_t gfp
, int cpu
)
3455 struct percpu_data
*pdata
= __percpu_disguise(__pdata
);
3456 int node
= cpu_to_node(cpu
);
3458 BUG_ON(pdata
->ptrs
[cpu
]);
3459 if (node_online(node
)) {
3460 /* FIXME: kzalloc_node(size, gfp, node) */
3461 pdata
->ptrs
[cpu
] = kmalloc_node(size
, gfp
, node
);
3462 if (pdata
->ptrs
[cpu
])
3463 memset(pdata
->ptrs
[cpu
], 0, size
);
3465 pdata
->ptrs
[cpu
] = kzalloc(size
, gfp
);
3466 return pdata
->ptrs
[cpu
];
3468 EXPORT_SYMBOL_GPL(percpu_populate
);
3471 * percpu_populate_mask - populate per-cpu data for more cpu's
3472 * @__pdata: per-cpu data to populate further
3473 * @size: size of per-cpu object
3474 * @gfp: may sleep or not etc.
3475 * @mask: populate per-cpu data for cpu's selected through mask bits
3477 * Per-cpu objects are populated with zeroed buffers.
3479 int __percpu_populate_mask(void *__pdata
, size_t size
, gfp_t gfp
,
3482 cpumask_t populated
= CPU_MASK_NONE
;
3485 for_each_cpu_mask(cpu
, *mask
)
3486 if (unlikely(!percpu_populate(__pdata
, size
, gfp
, cpu
))) {
3487 __percpu_depopulate_mask(__pdata
, &populated
);
3490 cpu_set(cpu
, populated
);
3493 EXPORT_SYMBOL_GPL(__percpu_populate_mask
);
3496 * percpu_alloc_mask - initial setup of per-cpu data
3497 * @size: size of per-cpu object
3498 * @gfp: may sleep or not etc.
3499 * @mask: populate per-data for cpu's selected through mask bits
3501 * Populating per-cpu data for all online cpu's would be a typical use case,
3502 * which is simplified by the percpu_alloc() wrapper.
3503 * Per-cpu objects are populated with zeroed buffers.
3505 void *__percpu_alloc_mask(size_t size
, gfp_t gfp
, cpumask_t
*mask
)
3507 void *pdata
= kzalloc(sizeof(struct percpu_data
), gfp
);
3508 void *__pdata
= __percpu_disguise(pdata
);
3510 if (unlikely(!pdata
))
3512 if (likely(!__percpu_populate_mask(__pdata
, size
, gfp
, mask
)))
3517 EXPORT_SYMBOL_GPL(__percpu_alloc_mask
);
3520 * percpu_free - final cleanup of per-cpu data
3521 * @__pdata: object to clean up
3523 * We simply clean up any per-cpu object left. No need for the client to
3524 * track and specify through a bis mask which per-cpu objects are to free.
3526 void percpu_free(void *__pdata
)
3528 __percpu_depopulate_mask(__pdata
, &cpu_possible_map
);
3529 kfree(__percpu_disguise(__pdata
));
3531 EXPORT_SYMBOL_GPL(percpu_free
);
3532 #endif /* CONFIG_SMP */
3535 * kmem_cache_free - Deallocate an object
3536 * @cachep: The cache the allocation was from.
3537 * @objp: The previously allocated object.
3539 * Free an object which was previously allocated from this
3542 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3544 unsigned long flags
;
3546 BUG_ON(virt_to_cache(objp
) != cachep
);
3548 local_irq_save(flags
);
3549 __cache_free(cachep
, objp
);
3550 local_irq_restore(flags
);
3552 EXPORT_SYMBOL(kmem_cache_free
);
3555 * kfree - free previously allocated memory
3556 * @objp: pointer returned by kmalloc.
3558 * If @objp is NULL, no operation is performed.
3560 * Don't free memory not originally allocated by kmalloc()
3561 * or you will run into trouble.
3563 void kfree(const void *objp
)
3565 struct kmem_cache
*c
;
3566 unsigned long flags
;
3568 if (unlikely(!objp
))
3570 local_irq_save(flags
);
3571 kfree_debugcheck(objp
);
3572 c
= virt_to_cache(objp
);
3573 debug_check_no_locks_freed(objp
, obj_size(c
));
3574 __cache_free(c
, (void *)objp
);
3575 local_irq_restore(flags
);
3577 EXPORT_SYMBOL(kfree
);
3579 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3581 return obj_size(cachep
);
3583 EXPORT_SYMBOL(kmem_cache_size
);
3585 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3587 return cachep
->name
;
3589 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3592 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3594 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3597 struct kmem_list3
*l3
;
3598 struct array_cache
*new_shared
;
3599 struct array_cache
**new_alien
;
3601 for_each_online_node(node
) {
3603 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3607 new_shared
= alloc_arraycache(node
,
3608 cachep
->shared
*cachep
->batchcount
,
3611 free_alien_cache(new_alien
);
3615 l3
= cachep
->nodelists
[node
];
3617 struct array_cache
*shared
= l3
->shared
;
3619 spin_lock_irq(&l3
->list_lock
);
3622 free_block(cachep
, shared
->entry
,
3623 shared
->avail
, node
);
3625 l3
->shared
= new_shared
;
3627 l3
->alien
= new_alien
;
3630 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3631 cachep
->batchcount
+ cachep
->num
;
3632 spin_unlock_irq(&l3
->list_lock
);
3634 free_alien_cache(new_alien
);
3637 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3639 free_alien_cache(new_alien
);
3644 kmem_list3_init(l3
);
3645 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3646 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3647 l3
->shared
= new_shared
;
3648 l3
->alien
= new_alien
;
3649 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3650 cachep
->batchcount
+ cachep
->num
;
3651 cachep
->nodelists
[node
] = l3
;
3656 if (!cachep
->next
.next
) {
3657 /* Cache is not active yet. Roll back what we did */
3660 if (cachep
->nodelists
[node
]) {
3661 l3
= cachep
->nodelists
[node
];
3664 free_alien_cache(l3
->alien
);
3666 cachep
->nodelists
[node
] = NULL
;
3674 struct ccupdate_struct
{
3675 struct kmem_cache
*cachep
;
3676 struct array_cache
*new[NR_CPUS
];
3679 static void do_ccupdate_local(void *info
)
3681 struct ccupdate_struct
*new = info
;
3682 struct array_cache
*old
;
3685 old
= cpu_cache_get(new->cachep
);
3687 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3688 new->new[smp_processor_id()] = old
;
3691 /* Always called with the cache_chain_mutex held */
3692 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3693 int batchcount
, int shared
)
3695 struct ccupdate_struct
new;
3698 memset(&new.new, 0, sizeof(new.new));
3699 for_each_online_cpu(i
) {
3700 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3703 for (i
--; i
>= 0; i
--)
3708 new.cachep
= cachep
;
3710 on_each_cpu(do_ccupdate_local
, (void *)&new, 1, 1);
3713 cachep
->batchcount
= batchcount
;
3714 cachep
->limit
= limit
;
3715 cachep
->shared
= shared
;
3717 for_each_online_cpu(i
) {
3718 struct array_cache
*ccold
= new.new[i
];
3721 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3722 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3723 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3727 err
= alloc_kmemlist(cachep
);
3729 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3730 cachep
->name
, -err
);
3736 /* Called with cache_chain_mutex held always */
3737 static void enable_cpucache(struct kmem_cache
*cachep
)
3743 * The head array serves three purposes:
3744 * - create a LIFO ordering, i.e. return objects that are cache-warm
3745 * - reduce the number of spinlock operations.
3746 * - reduce the number of linked list operations on the slab and
3747 * bufctl chains: array operations are cheaper.
3748 * The numbers are guessed, we should auto-tune as described by
3751 if (cachep
->buffer_size
> 131072)
3753 else if (cachep
->buffer_size
> PAGE_SIZE
)
3755 else if (cachep
->buffer_size
> 1024)
3757 else if (cachep
->buffer_size
> 256)
3763 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3764 * allocation behaviour: Most allocs on one cpu, most free operations
3765 * on another cpu. For these cases, an efficient object passing between
3766 * cpus is necessary. This is provided by a shared array. The array
3767 * replaces Bonwick's magazine layer.
3768 * On uniprocessor, it's functionally equivalent (but less efficient)
3769 * to a larger limit. Thus disabled by default.
3773 if (cachep
->buffer_size
<= PAGE_SIZE
)
3779 * With debugging enabled, large batchcount lead to excessively long
3780 * periods with disabled local interrupts. Limit the batchcount
3785 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3787 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3788 cachep
->name
, -err
);
3792 * Drain an array if it contains any elements taking the l3 lock only if
3793 * necessary. Note that the l3 listlock also protects the array_cache
3794 * if drain_array() is used on the shared array.
3796 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3797 struct array_cache
*ac
, int force
, int node
)
3801 if (!ac
|| !ac
->avail
)
3803 if (ac
->touched
&& !force
) {
3806 spin_lock_irq(&l3
->list_lock
);
3808 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3809 if (tofree
> ac
->avail
)
3810 tofree
= (ac
->avail
+ 1) / 2;
3811 free_block(cachep
, ac
->entry
, tofree
, node
);
3812 ac
->avail
-= tofree
;
3813 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3814 sizeof(void *) * ac
->avail
);
3816 spin_unlock_irq(&l3
->list_lock
);
3821 * cache_reap - Reclaim memory from caches.
3822 * @unused: unused parameter
3824 * Called from workqueue/eventd every few seconds.
3826 * - clear the per-cpu caches for this CPU.
3827 * - return freeable pages to the main free memory pool.
3829 * If we cannot acquire the cache chain mutex then just give up - we'll try
3830 * again on the next iteration.
3832 static void cache_reap(void *unused
)
3834 struct kmem_cache
*searchp
;
3835 struct kmem_list3
*l3
;
3836 int node
= numa_node_id();
3838 if (!mutex_trylock(&cache_chain_mutex
)) {
3839 /* Give up. Setup the next iteration. */
3840 schedule_delayed_work(&__get_cpu_var(reap_work
),
3845 list_for_each_entry(searchp
, &cache_chain
, next
) {
3849 * We only take the l3 lock if absolutely necessary and we
3850 * have established with reasonable certainty that
3851 * we can do some work if the lock was obtained.
3853 l3
= searchp
->nodelists
[node
];
3855 reap_alien(searchp
, l3
);
3857 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3860 * These are racy checks but it does not matter
3861 * if we skip one check or scan twice.
3863 if (time_after(l3
->next_reap
, jiffies
))
3866 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3868 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3870 if (l3
->free_touched
)
3871 l3
->free_touched
= 0;
3875 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
3876 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3877 STATS_ADD_REAPED(searchp
, freed
);
3883 mutex_unlock(&cache_chain_mutex
);
3885 refresh_cpu_vm_stats(smp_processor_id());
3886 /* Set up the next iteration */
3887 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3890 #ifdef CONFIG_PROC_FS
3892 static void print_slabinfo_header(struct seq_file
*m
)
3895 * Output format version, so at least we can change it
3896 * without _too_ many complaints.
3899 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3901 seq_puts(m
, "slabinfo - version: 2.1\n");
3903 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3904 "<objperslab> <pagesperslab>");
3905 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3906 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3908 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3909 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3910 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3915 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3918 struct list_head
*p
;
3920 mutex_lock(&cache_chain_mutex
);
3922 print_slabinfo_header(m
);
3923 p
= cache_chain
.next
;
3926 if (p
== &cache_chain
)
3929 return list_entry(p
, struct kmem_cache
, next
);
3932 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3934 struct kmem_cache
*cachep
= p
;
3936 return cachep
->next
.next
== &cache_chain
?
3937 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3940 static void s_stop(struct seq_file
*m
, void *p
)
3942 mutex_unlock(&cache_chain_mutex
);
3945 static int s_show(struct seq_file
*m
, void *p
)
3947 struct kmem_cache
*cachep
= p
;
3949 unsigned long active_objs
;
3950 unsigned long num_objs
;
3951 unsigned long active_slabs
= 0;
3952 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3956 struct kmem_list3
*l3
;
3960 for_each_online_node(node
) {
3961 l3
= cachep
->nodelists
[node
];
3966 spin_lock_irq(&l3
->list_lock
);
3968 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
3969 if (slabp
->inuse
!= cachep
->num
&& !error
)
3970 error
= "slabs_full accounting error";
3971 active_objs
+= cachep
->num
;
3974 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
3975 if (slabp
->inuse
== cachep
->num
&& !error
)
3976 error
= "slabs_partial inuse accounting error";
3977 if (!slabp
->inuse
&& !error
)
3978 error
= "slabs_partial/inuse accounting error";
3979 active_objs
+= slabp
->inuse
;
3982 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
3983 if (slabp
->inuse
&& !error
)
3984 error
= "slabs_free/inuse accounting error";
3987 free_objects
+= l3
->free_objects
;
3989 shared_avail
+= l3
->shared
->avail
;
3991 spin_unlock_irq(&l3
->list_lock
);
3993 num_slabs
+= active_slabs
;
3994 num_objs
= num_slabs
* cachep
->num
;
3995 if (num_objs
- active_objs
!= free_objects
&& !error
)
3996 error
= "free_objects accounting error";
3998 name
= cachep
->name
;
4000 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4002 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4003 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4004 cachep
->num
, (1 << cachep
->gfporder
));
4005 seq_printf(m
, " : tunables %4u %4u %4u",
4006 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4007 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4008 active_slabs
, num_slabs
, shared_avail
);
4011 unsigned long high
= cachep
->high_mark
;
4012 unsigned long allocs
= cachep
->num_allocations
;
4013 unsigned long grown
= cachep
->grown
;
4014 unsigned long reaped
= cachep
->reaped
;
4015 unsigned long errors
= cachep
->errors
;
4016 unsigned long max_freeable
= cachep
->max_freeable
;
4017 unsigned long node_allocs
= cachep
->node_allocs
;
4018 unsigned long node_frees
= cachep
->node_frees
;
4019 unsigned long overflows
= cachep
->node_overflow
;
4021 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4022 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4023 reaped
, errors
, max_freeable
, node_allocs
,
4024 node_frees
, overflows
);
4028 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4029 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4030 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4031 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4033 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4034 allochit
, allocmiss
, freehit
, freemiss
);
4042 * slabinfo_op - iterator that generates /proc/slabinfo
4051 * num-pages-per-slab
4052 * + further values on SMP and with statistics enabled
4055 struct seq_operations slabinfo_op
= {
4062 #define MAX_SLABINFO_WRITE 128
4064 * slabinfo_write - Tuning for the slab allocator
4066 * @buffer: user buffer
4067 * @count: data length
4070 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4071 size_t count
, loff_t
*ppos
)
4073 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4074 int limit
, batchcount
, shared
, res
;
4075 struct kmem_cache
*cachep
;
4077 if (count
> MAX_SLABINFO_WRITE
)
4079 if (copy_from_user(&kbuf
, buffer
, count
))
4081 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4083 tmp
= strchr(kbuf
, ' ');
4088 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4091 /* Find the cache in the chain of caches. */
4092 mutex_lock(&cache_chain_mutex
);
4094 list_for_each_entry(cachep
, &cache_chain
, next
) {
4095 if (!strcmp(cachep
->name
, kbuf
)) {
4096 if (limit
< 1 || batchcount
< 1 ||
4097 batchcount
> limit
|| shared
< 0) {
4100 res
= do_tune_cpucache(cachep
, limit
,
4101 batchcount
, shared
);
4106 mutex_unlock(&cache_chain_mutex
);
4112 #ifdef CONFIG_DEBUG_SLAB_LEAK
4114 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4117 struct list_head
*p
;
4119 mutex_lock(&cache_chain_mutex
);
4120 p
= cache_chain
.next
;
4123 if (p
== &cache_chain
)
4126 return list_entry(p
, struct kmem_cache
, next
);
4129 static inline int add_caller(unsigned long *n
, unsigned long v
)
4139 unsigned long *q
= p
+ 2 * i
;
4153 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4159 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4165 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4166 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4168 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4173 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4175 #ifdef CONFIG_KALLSYMS
4178 unsigned long offset
, size
;
4179 char namebuf
[KSYM_NAME_LEN
+1];
4181 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4184 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4186 seq_printf(m
, " [%s]", modname
);
4190 seq_printf(m
, "%p", (void *)address
);
4193 static int leaks_show(struct seq_file
*m
, void *p
)
4195 struct kmem_cache
*cachep
= p
;
4197 struct kmem_list3
*l3
;
4199 unsigned long *n
= m
->private;
4203 if (!(cachep
->flags
& SLAB_STORE_USER
))
4205 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4208 /* OK, we can do it */
4212 for_each_online_node(node
) {
4213 l3
= cachep
->nodelists
[node
];
4218 spin_lock_irq(&l3
->list_lock
);
4220 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4221 handle_slab(n
, cachep
, slabp
);
4222 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4223 handle_slab(n
, cachep
, slabp
);
4224 spin_unlock_irq(&l3
->list_lock
);
4226 name
= cachep
->name
;
4228 /* Increase the buffer size */
4229 mutex_unlock(&cache_chain_mutex
);
4230 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4232 /* Too bad, we are really out */
4234 mutex_lock(&cache_chain_mutex
);
4237 *(unsigned long *)m
->private = n
[0] * 2;
4239 mutex_lock(&cache_chain_mutex
);
4240 /* Now make sure this entry will be retried */
4244 for (i
= 0; i
< n
[1]; i
++) {
4245 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4246 show_symbol(m
, n
[2*i
+2]);
4252 struct seq_operations slabstats_op
= {
4253 .start
= leaks_start
,
4262 * ksize - get the actual amount of memory allocated for a given object
4263 * @objp: Pointer to the object
4265 * kmalloc may internally round up allocations and return more memory
4266 * than requested. ksize() can be used to determine the actual amount of
4267 * memory allocated. The caller may use this additional memory, even though
4268 * a smaller amount of memory was initially specified with the kmalloc call.
4269 * The caller must guarantee that objp points to a valid object previously
4270 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4271 * must not be freed during the duration of the call.
4273 unsigned int ksize(const void *objp
)
4275 if (unlikely(objp
== NULL
))
4278 return obj_size(virt_to_cache(objp
));