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 int 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 #define BAD_ALIEN_MAGIC 0x01020304ul
679 #ifdef CONFIG_LOCKDEP
682 * Slab sometimes uses the kmalloc slabs to store the slab headers
683 * for other slabs "off slab".
684 * The locking for this is tricky in that it nests within the locks
685 * of all other slabs in a few places; to deal with this special
686 * locking we put on-slab caches into a separate lock-class.
688 * We set lock class for alien array caches which are up during init.
689 * The lock annotation will be lost if all cpus of a node goes down and
690 * then comes back up during hotplug
692 static struct lock_class_key on_slab_l3_key
;
693 static struct lock_class_key on_slab_alc_key
;
695 static inline void init_lock_keys(void)
699 struct cache_sizes
*s
= malloc_sizes
;
701 while (s
->cs_size
!= ULONG_MAX
) {
703 struct array_cache
**alc
;
705 struct kmem_list3
*l3
= s
->cs_cachep
->nodelists
[q
];
706 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
708 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
711 * FIXME: This check for BAD_ALIEN_MAGIC
712 * should go away when common slab code is taught to
713 * work even without alien caches.
714 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
715 * for alloc_alien_cache,
717 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
721 lockdep_set_class(&alc
[r
]->lock
,
729 static inline void init_lock_keys(void)
734 /* Guard access to the cache-chain. */
735 static DEFINE_MUTEX(cache_chain_mutex
);
736 static struct list_head cache_chain
;
739 * chicken and egg problem: delay the per-cpu array allocation
740 * until the general caches are up.
750 * used by boot code to determine if it can use slab based allocator
752 int slab_is_available(void)
754 return g_cpucache_up
== FULL
;
757 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
759 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
761 return cachep
->array
[smp_processor_id()];
764 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
767 struct cache_sizes
*csizep
= malloc_sizes
;
770 /* This happens if someone tries to call
771 * kmem_cache_create(), or __kmalloc(), before
772 * the generic caches are initialized.
774 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
776 while (size
> csizep
->cs_size
)
780 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
781 * has cs_{dma,}cachep==NULL. Thus no special case
782 * for large kmalloc calls required.
784 if (unlikely(gfpflags
& GFP_DMA
))
785 return csizep
->cs_dmacachep
;
786 return csizep
->cs_cachep
;
789 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
791 return __find_general_cachep(size
, gfpflags
);
794 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
796 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
800 * Calculate the number of objects and left-over bytes for a given buffer size.
802 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
803 size_t align
, int flags
, size_t *left_over
,
808 size_t slab_size
= PAGE_SIZE
<< gfporder
;
811 * The slab management structure can be either off the slab or
812 * on it. For the latter case, the memory allocated for a
816 * - One kmem_bufctl_t for each object
817 * - Padding to respect alignment of @align
818 * - @buffer_size bytes for each object
820 * If the slab management structure is off the slab, then the
821 * alignment will already be calculated into the size. Because
822 * the slabs are all pages aligned, the objects will be at the
823 * correct alignment when allocated.
825 if (flags
& CFLGS_OFF_SLAB
) {
827 nr_objs
= slab_size
/ buffer_size
;
829 if (nr_objs
> SLAB_LIMIT
)
830 nr_objs
= SLAB_LIMIT
;
833 * Ignore padding for the initial guess. The padding
834 * is at most @align-1 bytes, and @buffer_size is at
835 * least @align. In the worst case, this result will
836 * be one greater than the number of objects that fit
837 * into the memory allocation when taking the padding
840 nr_objs
= (slab_size
- sizeof(struct slab
)) /
841 (buffer_size
+ sizeof(kmem_bufctl_t
));
844 * This calculated number will be either the right
845 * amount, or one greater than what we want.
847 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
851 if (nr_objs
> SLAB_LIMIT
)
852 nr_objs
= SLAB_LIMIT
;
854 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
857 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
860 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
862 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
865 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
866 function
, cachep
->name
, msg
);
872 * Special reaping functions for NUMA systems called from cache_reap().
873 * These take care of doing round robin flushing of alien caches (containing
874 * objects freed on different nodes from which they were allocated) and the
875 * flushing of remote pcps by calling drain_node_pages.
877 static DEFINE_PER_CPU(unsigned long, reap_node
);
879 static void init_reap_node(int cpu
)
883 node
= next_node(cpu_to_node(cpu
), node_online_map
);
884 if (node
== MAX_NUMNODES
)
885 node
= first_node(node_online_map
);
887 __get_cpu_var(reap_node
) = node
;
890 static void next_reap_node(void)
892 int node
= __get_cpu_var(reap_node
);
895 * Also drain per cpu pages on remote zones
897 if (node
!= numa_node_id())
898 drain_node_pages(node
);
900 node
= next_node(node
, node_online_map
);
901 if (unlikely(node
>= MAX_NUMNODES
))
902 node
= first_node(node_online_map
);
903 __get_cpu_var(reap_node
) = node
;
907 #define init_reap_node(cpu) do { } while (0)
908 #define next_reap_node(void) do { } while (0)
912 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
913 * via the workqueue/eventd.
914 * Add the CPU number into the expiration time to minimize the possibility of
915 * the CPUs getting into lockstep and contending for the global cache chain
918 static void __devinit
start_cpu_timer(int cpu
)
920 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
923 * When this gets called from do_initcalls via cpucache_init(),
924 * init_workqueues() has already run, so keventd will be setup
927 if (keventd_up() && reap_work
->func
== NULL
) {
929 INIT_WORK(reap_work
, cache_reap
, NULL
);
930 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
934 static struct array_cache
*alloc_arraycache(int node
, int entries
,
937 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
938 struct array_cache
*nc
= NULL
;
940 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
944 nc
->batchcount
= batchcount
;
946 spin_lock_init(&nc
->lock
);
952 * Transfer objects in one arraycache to another.
953 * Locking must be handled by the caller.
955 * Return the number of entries transferred.
957 static int transfer_objects(struct array_cache
*to
,
958 struct array_cache
*from
, unsigned int max
)
960 /* Figure out how many entries to transfer */
961 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
966 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
977 #define drain_alien_cache(cachep, alien) do { } while (0)
978 #define reap_alien(cachep, l3) do { } while (0)
980 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
982 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
985 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
989 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
994 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1000 static inline void *__cache_alloc_node(struct kmem_cache
*cachep
,
1001 gfp_t flags
, int nodeid
)
1006 #else /* CONFIG_NUMA */
1008 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1009 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1011 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
1013 struct array_cache
**ac_ptr
;
1014 int memsize
= sizeof(void *) * MAX_NUMNODES
;
1019 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1022 if (i
== node
|| !node_online(i
)) {
1026 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
1028 for (i
--; i
<= 0; i
--)
1038 static void free_alien_cache(struct array_cache
**ac_ptr
)
1049 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1050 struct array_cache
*ac
, int node
)
1052 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1055 spin_lock(&rl3
->list_lock
);
1057 * Stuff objects into the remote nodes shared array first.
1058 * That way we could avoid the overhead of putting the objects
1059 * into the free lists and getting them back later.
1062 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1064 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1066 spin_unlock(&rl3
->list_lock
);
1071 * Called from cache_reap() to regularly drain alien caches round robin.
1073 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1075 int node
= __get_cpu_var(reap_node
);
1078 struct array_cache
*ac
= l3
->alien
[node
];
1080 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1081 __drain_alien_cache(cachep
, ac
, node
);
1082 spin_unlock_irq(&ac
->lock
);
1087 static void drain_alien_cache(struct kmem_cache
*cachep
,
1088 struct array_cache
**alien
)
1091 struct array_cache
*ac
;
1092 unsigned long flags
;
1094 for_each_online_node(i
) {
1097 spin_lock_irqsave(&ac
->lock
, flags
);
1098 __drain_alien_cache(cachep
, ac
, i
);
1099 spin_unlock_irqrestore(&ac
->lock
, flags
);
1104 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1106 struct slab
*slabp
= virt_to_slab(objp
);
1107 int nodeid
= slabp
->nodeid
;
1108 struct kmem_list3
*l3
;
1109 struct array_cache
*alien
= NULL
;
1112 * Make sure we are not freeing a object from another node to the array
1113 * cache on this cpu.
1115 if (likely(slabp
->nodeid
== numa_node_id()))
1118 l3
= cachep
->nodelists
[numa_node_id()];
1119 STATS_INC_NODEFREES(cachep
);
1120 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1121 alien
= l3
->alien
[nodeid
];
1122 spin_lock(&alien
->lock
);
1123 if (unlikely(alien
->avail
== alien
->limit
)) {
1124 STATS_INC_ACOVERFLOW(cachep
);
1125 __drain_alien_cache(cachep
, alien
, nodeid
);
1127 alien
->entry
[alien
->avail
++] = objp
;
1128 spin_unlock(&alien
->lock
);
1130 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1131 free_block(cachep
, &objp
, 1, nodeid
);
1132 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1138 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1139 unsigned long action
, void *hcpu
)
1141 long cpu
= (long)hcpu
;
1142 struct kmem_cache
*cachep
;
1143 struct kmem_list3
*l3
= NULL
;
1144 int node
= cpu_to_node(cpu
);
1145 int memsize
= sizeof(struct kmem_list3
);
1148 case CPU_UP_PREPARE
:
1149 mutex_lock(&cache_chain_mutex
);
1151 * We need to do this right in the beginning since
1152 * alloc_arraycache's are going to use this list.
1153 * kmalloc_node allows us to add the slab to the right
1154 * kmem_list3 and not this cpu's kmem_list3
1157 list_for_each_entry(cachep
, &cache_chain
, next
) {
1159 * Set up the size64 kmemlist for cpu before we can
1160 * begin anything. Make sure some other cpu on this
1161 * node has not already allocated this
1163 if (!cachep
->nodelists
[node
]) {
1164 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1167 kmem_list3_init(l3
);
1168 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1169 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1172 * The l3s don't come and go as CPUs come and
1173 * go. cache_chain_mutex is sufficient
1176 cachep
->nodelists
[node
] = l3
;
1179 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1180 cachep
->nodelists
[node
]->free_limit
=
1181 (1 + nr_cpus_node(node
)) *
1182 cachep
->batchcount
+ cachep
->num
;
1183 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1187 * Now we can go ahead with allocating the shared arrays and
1190 list_for_each_entry(cachep
, &cache_chain
, next
) {
1191 struct array_cache
*nc
;
1192 struct array_cache
*shared
;
1193 struct array_cache
**alien
;
1195 nc
= alloc_arraycache(node
, cachep
->limit
,
1196 cachep
->batchcount
);
1199 shared
= alloc_arraycache(node
,
1200 cachep
->shared
* cachep
->batchcount
,
1205 alien
= alloc_alien_cache(node
, cachep
->limit
);
1208 cachep
->array
[cpu
] = nc
;
1209 l3
= cachep
->nodelists
[node
];
1212 spin_lock_irq(&l3
->list_lock
);
1215 * We are serialised from CPU_DEAD or
1216 * CPU_UP_CANCELLED by the cpucontrol lock
1218 l3
->shared
= shared
;
1227 spin_unlock_irq(&l3
->list_lock
);
1229 free_alien_cache(alien
);
1231 mutex_unlock(&cache_chain_mutex
);
1234 start_cpu_timer(cpu
);
1236 #ifdef CONFIG_HOTPLUG_CPU
1239 * Even if all the cpus of a node are down, we don't free the
1240 * kmem_list3 of any cache. This to avoid a race between
1241 * cpu_down, and a kmalloc allocation from another cpu for
1242 * memory from the node of the cpu going down. The list3
1243 * structure is usually allocated from kmem_cache_create() and
1244 * gets destroyed at kmem_cache_destroy().
1247 case CPU_UP_CANCELED
:
1248 mutex_lock(&cache_chain_mutex
);
1249 list_for_each_entry(cachep
, &cache_chain
, next
) {
1250 struct array_cache
*nc
;
1251 struct array_cache
*shared
;
1252 struct array_cache
**alien
;
1255 mask
= node_to_cpumask(node
);
1256 /* cpu is dead; no one can alloc from it. */
1257 nc
= cachep
->array
[cpu
];
1258 cachep
->array
[cpu
] = NULL
;
1259 l3
= cachep
->nodelists
[node
];
1262 goto free_array_cache
;
1264 spin_lock_irq(&l3
->list_lock
);
1266 /* Free limit for this kmem_list3 */
1267 l3
->free_limit
-= cachep
->batchcount
;
1269 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1271 if (!cpus_empty(mask
)) {
1272 spin_unlock_irq(&l3
->list_lock
);
1273 goto free_array_cache
;
1276 shared
= l3
->shared
;
1278 free_block(cachep
, l3
->shared
->entry
,
1279 l3
->shared
->avail
, node
);
1286 spin_unlock_irq(&l3
->list_lock
);
1290 drain_alien_cache(cachep
, alien
);
1291 free_alien_cache(alien
);
1297 * In the previous loop, all the objects were freed to
1298 * the respective cache's slabs, now we can go ahead and
1299 * shrink each nodelist to its limit.
1301 list_for_each_entry(cachep
, &cache_chain
, next
) {
1302 l3
= cachep
->nodelists
[node
];
1305 drain_freelist(cachep
, l3
, l3
->free_objects
);
1307 mutex_unlock(&cache_chain_mutex
);
1313 mutex_unlock(&cache_chain_mutex
);
1317 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1318 &cpuup_callback
, NULL
, 0
1322 * swap the static kmem_list3 with kmalloced memory
1324 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1327 struct kmem_list3
*ptr
;
1329 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1330 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1333 local_irq_disable();
1334 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1336 * Do not assume that spinlocks can be initialized via memcpy:
1338 spin_lock_init(&ptr
->list_lock
);
1340 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1341 cachep
->nodelists
[nodeid
] = ptr
;
1346 * Initialisation. Called after the page allocator have been initialised and
1347 * before smp_init().
1349 void __init
kmem_cache_init(void)
1352 struct cache_sizes
*sizes
;
1353 struct cache_names
*names
;
1357 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1358 kmem_list3_init(&initkmem_list3
[i
]);
1359 if (i
< MAX_NUMNODES
)
1360 cache_cache
.nodelists
[i
] = NULL
;
1364 * Fragmentation resistance on low memory - only use bigger
1365 * page orders on machines with more than 32MB of memory.
1367 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1368 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1370 /* Bootstrap is tricky, because several objects are allocated
1371 * from caches that do not exist yet:
1372 * 1) initialize the cache_cache cache: it contains the struct
1373 * kmem_cache structures of all caches, except cache_cache itself:
1374 * cache_cache is statically allocated.
1375 * Initially an __init data area is used for the head array and the
1376 * kmem_list3 structures, it's replaced with a kmalloc allocated
1377 * array at the end of the bootstrap.
1378 * 2) Create the first kmalloc cache.
1379 * The struct kmem_cache for the new cache is allocated normally.
1380 * An __init data area is used for the head array.
1381 * 3) Create the remaining kmalloc caches, with minimally sized
1383 * 4) Replace the __init data head arrays for cache_cache and the first
1384 * kmalloc cache with kmalloc allocated arrays.
1385 * 5) Replace the __init data for kmem_list3 for cache_cache and
1386 * the other cache's with kmalloc allocated memory.
1387 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1390 /* 1) create the cache_cache */
1391 INIT_LIST_HEAD(&cache_chain
);
1392 list_add(&cache_cache
.next
, &cache_chain
);
1393 cache_cache
.colour_off
= cache_line_size();
1394 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1395 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1397 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1400 for (order
= 0; order
< MAX_ORDER
; order
++) {
1401 cache_estimate(order
, cache_cache
.buffer_size
,
1402 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1403 if (cache_cache
.num
)
1406 BUG_ON(!cache_cache
.num
);
1407 cache_cache
.gfporder
= order
;
1408 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1409 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1410 sizeof(struct slab
), cache_line_size());
1412 /* 2+3) create the kmalloc caches */
1413 sizes
= malloc_sizes
;
1414 names
= cache_names
;
1417 * Initialize the caches that provide memory for the array cache and the
1418 * kmem_list3 structures first. Without this, further allocations will
1422 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1423 sizes
[INDEX_AC
].cs_size
,
1424 ARCH_KMALLOC_MINALIGN
,
1425 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1428 if (INDEX_AC
!= INDEX_L3
) {
1429 sizes
[INDEX_L3
].cs_cachep
=
1430 kmem_cache_create(names
[INDEX_L3
].name
,
1431 sizes
[INDEX_L3
].cs_size
,
1432 ARCH_KMALLOC_MINALIGN
,
1433 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1437 slab_early_init
= 0;
1439 while (sizes
->cs_size
!= ULONG_MAX
) {
1441 * For performance, all the general caches are L1 aligned.
1442 * This should be particularly beneficial on SMP boxes, as it
1443 * eliminates "false sharing".
1444 * Note for systems short on memory removing the alignment will
1445 * allow tighter packing of the smaller caches.
1447 if (!sizes
->cs_cachep
) {
1448 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1450 ARCH_KMALLOC_MINALIGN
,
1451 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1455 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1457 ARCH_KMALLOC_MINALIGN
,
1458 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1464 /* 4) Replace the bootstrap head arrays */
1466 struct array_cache
*ptr
;
1468 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1470 local_irq_disable();
1471 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1472 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1473 sizeof(struct arraycache_init
));
1475 * Do not assume that spinlocks can be initialized via memcpy:
1477 spin_lock_init(&ptr
->lock
);
1479 cache_cache
.array
[smp_processor_id()] = ptr
;
1482 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1484 local_irq_disable();
1485 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1486 != &initarray_generic
.cache
);
1487 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1488 sizeof(struct arraycache_init
));
1490 * Do not assume that spinlocks can be initialized via memcpy:
1492 spin_lock_init(&ptr
->lock
);
1494 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1498 /* 5) Replace the bootstrap kmem_list3's */
1501 /* Replace the static kmem_list3 structures for the boot cpu */
1502 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1505 for_each_online_node(node
) {
1506 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1507 &initkmem_list3
[SIZE_AC
+ node
], node
);
1509 if (INDEX_AC
!= INDEX_L3
) {
1510 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1511 &initkmem_list3
[SIZE_L3
+ node
],
1517 /* 6) resize the head arrays to their final sizes */
1519 struct kmem_cache
*cachep
;
1520 mutex_lock(&cache_chain_mutex
);
1521 list_for_each_entry(cachep
, &cache_chain
, next
)
1522 if (enable_cpucache(cachep
))
1524 mutex_unlock(&cache_chain_mutex
);
1527 /* Annotate slab for lockdep -- annotate the malloc caches */
1532 g_cpucache_up
= FULL
;
1535 * Register a cpu startup notifier callback that initializes
1536 * cpu_cache_get for all new cpus
1538 register_cpu_notifier(&cpucache_notifier
);
1541 * The reap timers are started later, with a module init call: That part
1542 * of the kernel is not yet operational.
1546 static int __init
cpucache_init(void)
1551 * Register the timers that return unneeded pages to the page allocator
1553 for_each_online_cpu(cpu
)
1554 start_cpu_timer(cpu
);
1557 __initcall(cpucache_init
);
1560 * Interface to system's page allocator. No need to hold the cache-lock.
1562 * If we requested dmaable memory, we will get it. Even if we
1563 * did not request dmaable memory, we might get it, but that
1564 * would be relatively rare and ignorable.
1566 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1574 * Nommu uses slab's for process anonymous memory allocations, and thus
1575 * requires __GFP_COMP to properly refcount higher order allocations
1577 flags
|= __GFP_COMP
;
1581 * Under NUMA we want memory on the indicated node. We will handle
1582 * the needed fallback ourselves since we want to serve from our
1583 * per node object lists first for other nodes.
1585 flags
|= cachep
->gfpflags
| GFP_THISNODE
;
1587 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1591 nr_pages
= (1 << cachep
->gfporder
);
1592 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1593 add_zone_page_state(page_zone(page
),
1594 NR_SLAB_RECLAIMABLE
, nr_pages
);
1596 add_zone_page_state(page_zone(page
),
1597 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1598 for (i
= 0; i
< nr_pages
; i
++)
1599 __SetPageSlab(page
+ i
);
1600 return page_address(page
);
1604 * Interface to system's page release.
1606 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1608 unsigned long i
= (1 << cachep
->gfporder
);
1609 struct page
*page
= virt_to_page(addr
);
1610 const unsigned long nr_freed
= i
;
1612 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1613 sub_zone_page_state(page_zone(page
),
1614 NR_SLAB_RECLAIMABLE
, nr_freed
);
1616 sub_zone_page_state(page_zone(page
),
1617 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1619 BUG_ON(!PageSlab(page
));
1620 __ClearPageSlab(page
);
1623 if (current
->reclaim_state
)
1624 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1625 free_pages((unsigned long)addr
, cachep
->gfporder
);
1628 static void kmem_rcu_free(struct rcu_head
*head
)
1630 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1631 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1633 kmem_freepages(cachep
, slab_rcu
->addr
);
1634 if (OFF_SLAB(cachep
))
1635 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1640 #ifdef CONFIG_DEBUG_PAGEALLOC
1641 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1642 unsigned long caller
)
1644 int size
= obj_size(cachep
);
1646 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1648 if (size
< 5 * sizeof(unsigned long))
1651 *addr
++ = 0x12345678;
1653 *addr
++ = smp_processor_id();
1654 size
-= 3 * sizeof(unsigned long);
1656 unsigned long *sptr
= &caller
;
1657 unsigned long svalue
;
1659 while (!kstack_end(sptr
)) {
1661 if (kernel_text_address(svalue
)) {
1663 size
-= sizeof(unsigned long);
1664 if (size
<= sizeof(unsigned long))
1670 *addr
++ = 0x87654321;
1674 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1676 int size
= obj_size(cachep
);
1677 addr
= &((char *)addr
)[obj_offset(cachep
)];
1679 memset(addr
, val
, size
);
1680 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1683 static void dump_line(char *data
, int offset
, int limit
)
1686 printk(KERN_ERR
"%03x:", offset
);
1687 for (i
= 0; i
< limit
; i
++)
1688 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1695 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1700 if (cachep
->flags
& SLAB_RED_ZONE
) {
1701 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1702 *dbg_redzone1(cachep
, objp
),
1703 *dbg_redzone2(cachep
, objp
));
1706 if (cachep
->flags
& SLAB_STORE_USER
) {
1707 printk(KERN_ERR
"Last user: [<%p>]",
1708 *dbg_userword(cachep
, objp
));
1709 print_symbol("(%s)",
1710 (unsigned long)*dbg_userword(cachep
, objp
));
1713 realobj
= (char *)objp
+ obj_offset(cachep
);
1714 size
= obj_size(cachep
);
1715 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1718 if (i
+ limit
> size
)
1720 dump_line(realobj
, i
, limit
);
1724 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1730 realobj
= (char *)objp
+ obj_offset(cachep
);
1731 size
= obj_size(cachep
);
1733 for (i
= 0; i
< size
; i
++) {
1734 char exp
= POISON_FREE
;
1737 if (realobj
[i
] != exp
) {
1743 "Slab corruption: start=%p, len=%d\n",
1745 print_objinfo(cachep
, objp
, 0);
1747 /* Hexdump the affected line */
1750 if (i
+ limit
> size
)
1752 dump_line(realobj
, i
, limit
);
1755 /* Limit to 5 lines */
1761 /* Print some data about the neighboring objects, if they
1764 struct slab
*slabp
= virt_to_slab(objp
);
1767 objnr
= obj_to_index(cachep
, slabp
, objp
);
1769 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1770 realobj
= (char *)objp
+ obj_offset(cachep
);
1771 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1773 print_objinfo(cachep
, objp
, 2);
1775 if (objnr
+ 1 < cachep
->num
) {
1776 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1777 realobj
= (char *)objp
+ obj_offset(cachep
);
1778 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1780 print_objinfo(cachep
, objp
, 2);
1788 * slab_destroy_objs - destroy a slab and its objects
1789 * @cachep: cache pointer being destroyed
1790 * @slabp: slab pointer being destroyed
1792 * Call the registered destructor for each object in a slab that is being
1795 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1798 for (i
= 0; i
< cachep
->num
; i
++) {
1799 void *objp
= index_to_obj(cachep
, slabp
, i
);
1801 if (cachep
->flags
& SLAB_POISON
) {
1802 #ifdef CONFIG_DEBUG_PAGEALLOC
1803 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1805 kernel_map_pages(virt_to_page(objp
),
1806 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1808 check_poison_obj(cachep
, objp
);
1810 check_poison_obj(cachep
, objp
);
1813 if (cachep
->flags
& SLAB_RED_ZONE
) {
1814 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1815 slab_error(cachep
, "start of a freed object "
1817 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1818 slab_error(cachep
, "end of a freed object "
1821 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1822 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1826 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1830 for (i
= 0; i
< cachep
->num
; i
++) {
1831 void *objp
= index_to_obj(cachep
, slabp
, i
);
1832 (cachep
->dtor
) (objp
, cachep
, 0);
1839 * slab_destroy - destroy and release all objects in a slab
1840 * @cachep: cache pointer being destroyed
1841 * @slabp: slab pointer being destroyed
1843 * Destroy all the objs in a slab, and release the mem back to the system.
1844 * Before calling the slab must have been unlinked from the cache. The
1845 * cache-lock is not held/needed.
1847 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1849 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1851 slab_destroy_objs(cachep
, slabp
);
1852 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1853 struct slab_rcu
*slab_rcu
;
1855 slab_rcu
= (struct slab_rcu
*)slabp
;
1856 slab_rcu
->cachep
= cachep
;
1857 slab_rcu
->addr
= addr
;
1858 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1860 kmem_freepages(cachep
, addr
);
1861 if (OFF_SLAB(cachep
))
1862 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1867 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1868 * size of kmem_list3.
1870 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1874 for_each_online_node(node
) {
1875 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1876 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1878 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1882 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1885 struct kmem_list3
*l3
;
1887 for_each_online_cpu(i
)
1888 kfree(cachep
->array
[i
]);
1890 /* NUMA: free the list3 structures */
1891 for_each_online_node(i
) {
1892 l3
= cachep
->nodelists
[i
];
1895 free_alien_cache(l3
->alien
);
1899 kmem_cache_free(&cache_cache
, cachep
);
1904 * calculate_slab_order - calculate size (page order) of slabs
1905 * @cachep: pointer to the cache that is being created
1906 * @size: size of objects to be created in this cache.
1907 * @align: required alignment for the objects.
1908 * @flags: slab allocation flags
1910 * Also calculates the number of objects per slab.
1912 * This could be made much more intelligent. For now, try to avoid using
1913 * high order pages for slabs. When the gfp() functions are more friendly
1914 * towards high-order requests, this should be changed.
1916 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1917 size_t size
, size_t align
, unsigned long flags
)
1919 unsigned long offslab_limit
;
1920 size_t left_over
= 0;
1923 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1927 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1931 if (flags
& CFLGS_OFF_SLAB
) {
1933 * Max number of objs-per-slab for caches which
1934 * use off-slab slabs. Needed to avoid a possible
1935 * looping condition in cache_grow().
1937 offslab_limit
= size
- sizeof(struct slab
);
1938 offslab_limit
/= sizeof(kmem_bufctl_t
);
1940 if (num
> offslab_limit
)
1944 /* Found something acceptable - save it away */
1946 cachep
->gfporder
= gfporder
;
1947 left_over
= remainder
;
1950 * A VFS-reclaimable slab tends to have most allocations
1951 * as GFP_NOFS and we really don't want to have to be allocating
1952 * higher-order pages when we are unable to shrink dcache.
1954 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1958 * Large number of objects is good, but very large slabs are
1959 * currently bad for the gfp()s.
1961 if (gfporder
>= slab_break_gfp_order
)
1965 * Acceptable internal fragmentation?
1967 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1973 static int setup_cpu_cache(struct kmem_cache
*cachep
)
1975 if (g_cpucache_up
== FULL
)
1976 return enable_cpucache(cachep
);
1978 if (g_cpucache_up
== NONE
) {
1980 * Note: the first kmem_cache_create must create the cache
1981 * that's used by kmalloc(24), otherwise the creation of
1982 * further caches will BUG().
1984 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1987 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1988 * the first cache, then we need to set up all its list3s,
1989 * otherwise the creation of further caches will BUG().
1991 set_up_list3s(cachep
, SIZE_AC
);
1992 if (INDEX_AC
== INDEX_L3
)
1993 g_cpucache_up
= PARTIAL_L3
;
1995 g_cpucache_up
= PARTIAL_AC
;
1997 cachep
->array
[smp_processor_id()] =
1998 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
2000 if (g_cpucache_up
== PARTIAL_AC
) {
2001 set_up_list3s(cachep
, SIZE_L3
);
2002 g_cpucache_up
= PARTIAL_L3
;
2005 for_each_online_node(node
) {
2006 cachep
->nodelists
[node
] =
2007 kmalloc_node(sizeof(struct kmem_list3
),
2009 BUG_ON(!cachep
->nodelists
[node
]);
2010 kmem_list3_init(cachep
->nodelists
[node
]);
2014 cachep
->nodelists
[numa_node_id()]->next_reap
=
2015 jiffies
+ REAPTIMEOUT_LIST3
+
2016 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2018 cpu_cache_get(cachep
)->avail
= 0;
2019 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2020 cpu_cache_get(cachep
)->batchcount
= 1;
2021 cpu_cache_get(cachep
)->touched
= 0;
2022 cachep
->batchcount
= 1;
2023 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2028 * kmem_cache_create - Create a cache.
2029 * @name: A string which is used in /proc/slabinfo to identify this cache.
2030 * @size: The size of objects to be created in this cache.
2031 * @align: The required alignment for the objects.
2032 * @flags: SLAB flags
2033 * @ctor: A constructor for the objects.
2034 * @dtor: A destructor for the objects.
2036 * Returns a ptr to the cache on success, NULL on failure.
2037 * Cannot be called within a int, but can be interrupted.
2038 * The @ctor is run when new pages are allocated by the cache
2039 * and the @dtor is run before the pages are handed back.
2041 * @name must be valid until the cache is destroyed. This implies that
2042 * the module calling this has to destroy the cache before getting unloaded.
2046 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2047 * to catch references to uninitialised memory.
2049 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2050 * for buffer overruns.
2052 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2053 * cacheline. This can be beneficial if you're counting cycles as closely
2057 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2058 unsigned long flags
,
2059 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
2060 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
2062 size_t left_over
, slab_size
, ralign
;
2063 struct kmem_cache
*cachep
= NULL
, *pc
;
2066 * Sanity checks... these are all serious usage bugs.
2068 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2069 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
2070 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
2076 * Prevent CPUs from coming and going.
2077 * lock_cpu_hotplug() nests outside cache_chain_mutex
2081 mutex_lock(&cache_chain_mutex
);
2083 list_for_each_entry(pc
, &cache_chain
, next
) {
2084 mm_segment_t old_fs
= get_fs();
2089 * This happens when the module gets unloaded and doesn't
2090 * destroy its slab cache and no-one else reuses the vmalloc
2091 * area of the module. Print a warning.
2094 res
= __get_user(tmp
, pc
->name
);
2097 printk("SLAB: cache with size %d has lost its name\n",
2102 if (!strcmp(pc
->name
, name
)) {
2103 printk("kmem_cache_create: duplicate cache %s\n", name
);
2110 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2111 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
2112 /* No constructor, but inital state check requested */
2113 printk(KERN_ERR
"%s: No con, but init state check "
2114 "requested - %s\n", __FUNCTION__
, name
);
2115 flags
&= ~SLAB_DEBUG_INITIAL
;
2119 * Enable redzoning and last user accounting, except for caches with
2120 * large objects, if the increased size would increase the object size
2121 * above the next power of two: caches with object sizes just above a
2122 * power of two have a significant amount of internal fragmentation.
2124 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2125 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2126 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2127 flags
|= SLAB_POISON
;
2129 if (flags
& SLAB_DESTROY_BY_RCU
)
2130 BUG_ON(flags
& SLAB_POISON
);
2132 if (flags
& SLAB_DESTROY_BY_RCU
)
2136 * Always checks flags, a caller might be expecting debug support which
2139 BUG_ON(flags
& ~CREATE_MASK
);
2142 * Check that size is in terms of words. This is needed to avoid
2143 * unaligned accesses for some archs when redzoning is used, and makes
2144 * sure any on-slab bufctl's are also correctly aligned.
2146 if (size
& (BYTES_PER_WORD
- 1)) {
2147 size
+= (BYTES_PER_WORD
- 1);
2148 size
&= ~(BYTES_PER_WORD
- 1);
2151 /* calculate the final buffer alignment: */
2153 /* 1) arch recommendation: can be overridden for debug */
2154 if (flags
& SLAB_HWCACHE_ALIGN
) {
2156 * Default alignment: as specified by the arch code. Except if
2157 * an object is really small, then squeeze multiple objects into
2160 ralign
= cache_line_size();
2161 while (size
<= ralign
/ 2)
2164 ralign
= BYTES_PER_WORD
;
2168 * Redzoning and user store require word alignment. Note this will be
2169 * overridden by architecture or caller mandated alignment if either
2170 * is greater than BYTES_PER_WORD.
2172 if (flags
& SLAB_RED_ZONE
|| flags
& SLAB_STORE_USER
)
2173 ralign
= BYTES_PER_WORD
;
2175 /* 2) arch mandated alignment: disables debug if necessary */
2176 if (ralign
< ARCH_SLAB_MINALIGN
) {
2177 ralign
= ARCH_SLAB_MINALIGN
;
2178 if (ralign
> BYTES_PER_WORD
)
2179 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2181 /* 3) caller mandated alignment: disables debug if necessary */
2182 if (ralign
< align
) {
2184 if (ralign
> BYTES_PER_WORD
)
2185 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2192 /* Get cache's description obj. */
2193 cachep
= kmem_cache_zalloc(&cache_cache
, SLAB_KERNEL
);
2198 cachep
->obj_size
= size
;
2201 * Both debugging options require word-alignment which is calculated
2204 if (flags
& SLAB_RED_ZONE
) {
2205 /* add space for red zone words */
2206 cachep
->obj_offset
+= BYTES_PER_WORD
;
2207 size
+= 2 * BYTES_PER_WORD
;
2209 if (flags
& SLAB_STORE_USER
) {
2210 /* user store requires one word storage behind the end of
2213 size
+= BYTES_PER_WORD
;
2215 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2216 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2217 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2218 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2225 * Determine if the slab management is 'on' or 'off' slab.
2226 * (bootstrapping cannot cope with offslab caches so don't do
2229 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2231 * Size is large, assume best to place the slab management obj
2232 * off-slab (should allow better packing of objs).
2234 flags
|= CFLGS_OFF_SLAB
;
2236 size
= ALIGN(size
, align
);
2238 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2241 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2242 kmem_cache_free(&cache_cache
, cachep
);
2246 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2247 + sizeof(struct slab
), align
);
2250 * If the slab has been placed off-slab, and we have enough space then
2251 * move it on-slab. This is at the expense of any extra colouring.
2253 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2254 flags
&= ~CFLGS_OFF_SLAB
;
2255 left_over
-= slab_size
;
2258 if (flags
& CFLGS_OFF_SLAB
) {
2259 /* really off slab. No need for manual alignment */
2261 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2264 cachep
->colour_off
= cache_line_size();
2265 /* Offset must be a multiple of the alignment. */
2266 if (cachep
->colour_off
< align
)
2267 cachep
->colour_off
= align
;
2268 cachep
->colour
= left_over
/ cachep
->colour_off
;
2269 cachep
->slab_size
= slab_size
;
2270 cachep
->flags
= flags
;
2271 cachep
->gfpflags
= 0;
2272 if (flags
& SLAB_CACHE_DMA
)
2273 cachep
->gfpflags
|= GFP_DMA
;
2274 cachep
->buffer_size
= size
;
2276 if (flags
& CFLGS_OFF_SLAB
) {
2277 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2279 * This is a possibility for one of the malloc_sizes caches.
2280 * But since we go off slab only for object size greater than
2281 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2282 * this should not happen at all.
2283 * But leave a BUG_ON for some lucky dude.
2285 BUG_ON(!cachep
->slabp_cache
);
2287 cachep
->ctor
= ctor
;
2288 cachep
->dtor
= dtor
;
2289 cachep
->name
= name
;
2291 if (setup_cpu_cache(cachep
)) {
2292 __kmem_cache_destroy(cachep
);
2297 /* cache setup completed, link it into the list */
2298 list_add(&cachep
->next
, &cache_chain
);
2300 if (!cachep
&& (flags
& SLAB_PANIC
))
2301 panic("kmem_cache_create(): failed to create slab `%s'\n",
2303 mutex_unlock(&cache_chain_mutex
);
2304 unlock_cpu_hotplug();
2307 EXPORT_SYMBOL(kmem_cache_create
);
2310 static void check_irq_off(void)
2312 BUG_ON(!irqs_disabled());
2315 static void check_irq_on(void)
2317 BUG_ON(irqs_disabled());
2320 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2324 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2328 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2332 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2337 #define check_irq_off() do { } while(0)
2338 #define check_irq_on() do { } while(0)
2339 #define check_spinlock_acquired(x) do { } while(0)
2340 #define check_spinlock_acquired_node(x, y) do { } while(0)
2343 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2344 struct array_cache
*ac
,
2345 int force
, int node
);
2347 static void do_drain(void *arg
)
2349 struct kmem_cache
*cachep
= arg
;
2350 struct array_cache
*ac
;
2351 int node
= numa_node_id();
2354 ac
= cpu_cache_get(cachep
);
2355 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2356 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2357 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2361 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2363 struct kmem_list3
*l3
;
2366 on_each_cpu(do_drain
, cachep
, 1, 1);
2368 for_each_online_node(node
) {
2369 l3
= cachep
->nodelists
[node
];
2370 if (l3
&& l3
->alien
)
2371 drain_alien_cache(cachep
, l3
->alien
);
2374 for_each_online_node(node
) {
2375 l3
= cachep
->nodelists
[node
];
2377 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2382 * Remove slabs from the list of free slabs.
2383 * Specify the number of slabs to drain in tofree.
2385 * Returns the actual number of slabs released.
2387 static int drain_freelist(struct kmem_cache
*cache
,
2388 struct kmem_list3
*l3
, int tofree
)
2390 struct list_head
*p
;
2395 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2397 spin_lock_irq(&l3
->list_lock
);
2398 p
= l3
->slabs_free
.prev
;
2399 if (p
== &l3
->slabs_free
) {
2400 spin_unlock_irq(&l3
->list_lock
);
2404 slabp
= list_entry(p
, struct slab
, list
);
2406 BUG_ON(slabp
->inuse
);
2408 list_del(&slabp
->list
);
2410 * Safe to drop the lock. The slab is no longer linked
2413 l3
->free_objects
-= cache
->num
;
2414 spin_unlock_irq(&l3
->list_lock
);
2415 slab_destroy(cache
, slabp
);
2422 static int __cache_shrink(struct kmem_cache
*cachep
)
2425 struct kmem_list3
*l3
;
2427 drain_cpu_caches(cachep
);
2430 for_each_online_node(i
) {
2431 l3
= cachep
->nodelists
[i
];
2435 drain_freelist(cachep
, l3
, l3
->free_objects
);
2437 ret
+= !list_empty(&l3
->slabs_full
) ||
2438 !list_empty(&l3
->slabs_partial
);
2440 return (ret
? 1 : 0);
2444 * kmem_cache_shrink - Shrink a cache.
2445 * @cachep: The cache to shrink.
2447 * Releases as many slabs as possible for a cache.
2448 * To help debugging, a zero exit status indicates all slabs were released.
2450 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2452 BUG_ON(!cachep
|| in_interrupt());
2454 return __cache_shrink(cachep
);
2456 EXPORT_SYMBOL(kmem_cache_shrink
);
2459 * kmem_cache_destroy - delete a cache
2460 * @cachep: the cache to destroy
2462 * Remove a struct kmem_cache object from the slab cache.
2464 * It is expected this function will be called by a module when it is
2465 * unloaded. This will remove the cache completely, and avoid a duplicate
2466 * cache being allocated each time a module is loaded and unloaded, if the
2467 * module doesn't have persistent in-kernel storage across loads and unloads.
2469 * The cache must be empty before calling this function.
2471 * The caller must guarantee that noone will allocate memory from the cache
2472 * during the kmem_cache_destroy().
2474 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2476 BUG_ON(!cachep
|| in_interrupt());
2478 /* Don't let CPUs to come and go */
2481 /* Find the cache in the chain of caches. */
2482 mutex_lock(&cache_chain_mutex
);
2484 * the chain is never empty, cache_cache is never destroyed
2486 list_del(&cachep
->next
);
2487 mutex_unlock(&cache_chain_mutex
);
2489 if (__cache_shrink(cachep
)) {
2490 slab_error(cachep
, "Can't free all objects");
2491 mutex_lock(&cache_chain_mutex
);
2492 list_add(&cachep
->next
, &cache_chain
);
2493 mutex_unlock(&cache_chain_mutex
);
2494 unlock_cpu_hotplug();
2498 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2501 __kmem_cache_destroy(cachep
);
2502 unlock_cpu_hotplug();
2504 EXPORT_SYMBOL(kmem_cache_destroy
);
2507 * Get the memory for a slab management obj.
2508 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2509 * always come from malloc_sizes caches. The slab descriptor cannot
2510 * come from the same cache which is getting created because,
2511 * when we are searching for an appropriate cache for these
2512 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2513 * If we are creating a malloc_sizes cache here it would not be visible to
2514 * kmem_find_general_cachep till the initialization is complete.
2515 * Hence we cannot have slabp_cache same as the original cache.
2517 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2518 int colour_off
, gfp_t local_flags
,
2523 if (OFF_SLAB(cachep
)) {
2524 /* Slab management obj is off-slab. */
2525 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2526 local_flags
, nodeid
);
2530 slabp
= objp
+ colour_off
;
2531 colour_off
+= cachep
->slab_size
;
2534 slabp
->colouroff
= colour_off
;
2535 slabp
->s_mem
= objp
+ colour_off
;
2536 slabp
->nodeid
= nodeid
;
2540 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2542 return (kmem_bufctl_t
*) (slabp
+ 1);
2545 static void cache_init_objs(struct kmem_cache
*cachep
,
2546 struct slab
*slabp
, unsigned long ctor_flags
)
2550 for (i
= 0; i
< cachep
->num
; i
++) {
2551 void *objp
= index_to_obj(cachep
, slabp
, i
);
2553 /* need to poison the objs? */
2554 if (cachep
->flags
& SLAB_POISON
)
2555 poison_obj(cachep
, objp
, POISON_FREE
);
2556 if (cachep
->flags
& SLAB_STORE_USER
)
2557 *dbg_userword(cachep
, objp
) = NULL
;
2559 if (cachep
->flags
& SLAB_RED_ZONE
) {
2560 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2561 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2564 * Constructors are not allowed to allocate memory from the same
2565 * cache which they are a constructor for. Otherwise, deadlock.
2566 * They must also be threaded.
2568 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2569 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2572 if (cachep
->flags
& SLAB_RED_ZONE
) {
2573 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2574 slab_error(cachep
, "constructor overwrote the"
2575 " end of an object");
2576 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2577 slab_error(cachep
, "constructor overwrote the"
2578 " start of an object");
2580 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2581 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2582 kernel_map_pages(virt_to_page(objp
),
2583 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2586 cachep
->ctor(objp
, cachep
, ctor_flags
);
2588 slab_bufctl(slabp
)[i
] = i
+ 1;
2590 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2594 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2596 if (flags
& SLAB_DMA
)
2597 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2599 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2602 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2605 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2609 next
= slab_bufctl(slabp
)[slabp
->free
];
2611 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2612 WARN_ON(slabp
->nodeid
!= nodeid
);
2619 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2620 void *objp
, int nodeid
)
2622 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2625 /* Verify that the slab belongs to the intended node */
2626 WARN_ON(slabp
->nodeid
!= nodeid
);
2628 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2629 printk(KERN_ERR
"slab: double free detected in cache "
2630 "'%s', objp %p\n", cachep
->name
, objp
);
2634 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2635 slabp
->free
= objnr
;
2640 * Map pages beginning at addr to the given cache and slab. This is required
2641 * for the slab allocator to be able to lookup the cache and slab of a
2642 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2644 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2650 page
= virt_to_page(addr
);
2653 if (likely(!PageCompound(page
)))
2654 nr_pages
<<= cache
->gfporder
;
2657 page_set_cache(page
, cache
);
2658 page_set_slab(page
, slab
);
2660 } while (--nr_pages
);
2664 * Grow (by 1) the number of slabs within a cache. This is called by
2665 * kmem_cache_alloc() when there are no active objs left in a cache.
2667 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2673 unsigned long ctor_flags
;
2674 struct kmem_list3
*l3
;
2677 * Be lazy and only check for valid flags here, keeping it out of the
2678 * critical path in kmem_cache_alloc().
2680 BUG_ON(flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
));
2681 if (flags
& SLAB_NO_GROW
)
2684 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2685 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2686 if (!(local_flags
& __GFP_WAIT
))
2688 * Not allowed to sleep. Need to tell a constructor about
2689 * this - it might need to know...
2691 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2693 /* Take the l3 list lock to change the colour_next on this node */
2695 l3
= cachep
->nodelists
[nodeid
];
2696 spin_lock(&l3
->list_lock
);
2698 /* Get colour for the slab, and cal the next value. */
2699 offset
= l3
->colour_next
;
2701 if (l3
->colour_next
>= cachep
->colour
)
2702 l3
->colour_next
= 0;
2703 spin_unlock(&l3
->list_lock
);
2705 offset
*= cachep
->colour_off
;
2707 if (local_flags
& __GFP_WAIT
)
2711 * The test for missing atomic flag is performed here, rather than
2712 * the more obvious place, simply to reduce the critical path length
2713 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2714 * will eventually be caught here (where it matters).
2716 kmem_flagcheck(cachep
, flags
);
2719 * Get mem for the objs. Attempt to allocate a physical page from
2722 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2726 /* Get slab management. */
2727 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
, nodeid
);
2731 slabp
->nodeid
= nodeid
;
2732 slab_map_pages(cachep
, slabp
, objp
);
2734 cache_init_objs(cachep
, slabp
, ctor_flags
);
2736 if (local_flags
& __GFP_WAIT
)
2737 local_irq_disable();
2739 spin_lock(&l3
->list_lock
);
2741 /* Make slab active. */
2742 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2743 STATS_INC_GROWN(cachep
);
2744 l3
->free_objects
+= cachep
->num
;
2745 spin_unlock(&l3
->list_lock
);
2748 kmem_freepages(cachep
, objp
);
2750 if (local_flags
& __GFP_WAIT
)
2751 local_irq_disable();
2758 * Perform extra freeing checks:
2759 * - detect bad pointers.
2760 * - POISON/RED_ZONE checking
2761 * - destructor calls, for caches with POISON+dtor
2763 static void kfree_debugcheck(const void *objp
)
2767 if (!virt_addr_valid(objp
)) {
2768 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2769 (unsigned long)objp
);
2772 page
= virt_to_page(objp
);
2773 if (!PageSlab(page
)) {
2774 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2775 (unsigned long)objp
);
2780 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2782 unsigned long redzone1
, redzone2
;
2784 redzone1
= *dbg_redzone1(cache
, obj
);
2785 redzone2
= *dbg_redzone2(cache
, obj
);
2790 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2793 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2794 slab_error(cache
, "double free detected");
2796 slab_error(cache
, "memory outside object was overwritten");
2798 printk(KERN_ERR
"%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2799 obj
, redzone1
, redzone2
);
2802 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2809 objp
-= obj_offset(cachep
);
2810 kfree_debugcheck(objp
);
2811 page
= virt_to_page(objp
);
2813 slabp
= page_get_slab(page
);
2815 if (cachep
->flags
& SLAB_RED_ZONE
) {
2816 verify_redzone_free(cachep
, objp
);
2817 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2818 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2820 if (cachep
->flags
& SLAB_STORE_USER
)
2821 *dbg_userword(cachep
, objp
) = caller
;
2823 objnr
= obj_to_index(cachep
, slabp
, objp
);
2825 BUG_ON(objnr
>= cachep
->num
);
2826 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2828 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2830 * Need to call the slab's constructor so the caller can
2831 * perform a verify of its state (debugging). Called without
2832 * the cache-lock held.
2834 cachep
->ctor(objp
+ obj_offset(cachep
),
2835 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2837 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2838 /* we want to cache poison the object,
2839 * call the destruction callback
2841 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2843 #ifdef CONFIG_DEBUG_SLAB_LEAK
2844 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2846 if (cachep
->flags
& SLAB_POISON
) {
2847 #ifdef CONFIG_DEBUG_PAGEALLOC
2848 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2849 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2850 kernel_map_pages(virt_to_page(objp
),
2851 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2853 poison_obj(cachep
, objp
, POISON_FREE
);
2856 poison_obj(cachep
, objp
, POISON_FREE
);
2862 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2867 /* Check slab's freelist to see if this obj is there. */
2868 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2870 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2873 if (entries
!= cachep
->num
- slabp
->inuse
) {
2875 printk(KERN_ERR
"slab: Internal list corruption detected in "
2876 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2877 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2879 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2882 printk("\n%03x:", i
);
2883 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2890 #define kfree_debugcheck(x) do { } while(0)
2891 #define cache_free_debugcheck(x,objp,z) (objp)
2892 #define check_slabp(x,y) do { } while(0)
2895 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2898 struct kmem_list3
*l3
;
2899 struct array_cache
*ac
;
2902 ac
= cpu_cache_get(cachep
);
2904 batchcount
= ac
->batchcount
;
2905 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2907 * If there was little recent activity on this cache, then
2908 * perform only a partial refill. Otherwise we could generate
2911 batchcount
= BATCHREFILL_LIMIT
;
2913 l3
= cachep
->nodelists
[numa_node_id()];
2915 BUG_ON(ac
->avail
> 0 || !l3
);
2916 spin_lock(&l3
->list_lock
);
2918 /* See if we can refill from the shared array */
2919 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2922 while (batchcount
> 0) {
2923 struct list_head
*entry
;
2925 /* Get slab alloc is to come from. */
2926 entry
= l3
->slabs_partial
.next
;
2927 if (entry
== &l3
->slabs_partial
) {
2928 l3
->free_touched
= 1;
2929 entry
= l3
->slabs_free
.next
;
2930 if (entry
== &l3
->slabs_free
)
2934 slabp
= list_entry(entry
, struct slab
, list
);
2935 check_slabp(cachep
, slabp
);
2936 check_spinlock_acquired(cachep
);
2937 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2938 STATS_INC_ALLOCED(cachep
);
2939 STATS_INC_ACTIVE(cachep
);
2940 STATS_SET_HIGH(cachep
);
2942 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2945 check_slabp(cachep
, slabp
);
2947 /* move slabp to correct slabp list: */
2948 list_del(&slabp
->list
);
2949 if (slabp
->free
== BUFCTL_END
)
2950 list_add(&slabp
->list
, &l3
->slabs_full
);
2952 list_add(&slabp
->list
, &l3
->slabs_partial
);
2956 l3
->free_objects
-= ac
->avail
;
2958 spin_unlock(&l3
->list_lock
);
2960 if (unlikely(!ac
->avail
)) {
2962 x
= cache_grow(cachep
, flags
, numa_node_id());
2964 /* cache_grow can reenable interrupts, then ac could change. */
2965 ac
= cpu_cache_get(cachep
);
2966 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2969 if (!ac
->avail
) /* objects refilled by interrupt? */
2973 return ac
->entry
[--ac
->avail
];
2976 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2979 might_sleep_if(flags
& __GFP_WAIT
);
2981 kmem_flagcheck(cachep
, flags
);
2986 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2987 gfp_t flags
, void *objp
, void *caller
)
2991 if (cachep
->flags
& SLAB_POISON
) {
2992 #ifdef CONFIG_DEBUG_PAGEALLOC
2993 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2994 kernel_map_pages(virt_to_page(objp
),
2995 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2997 check_poison_obj(cachep
, objp
);
2999 check_poison_obj(cachep
, objp
);
3001 poison_obj(cachep
, objp
, POISON_INUSE
);
3003 if (cachep
->flags
& SLAB_STORE_USER
)
3004 *dbg_userword(cachep
, objp
) = caller
;
3006 if (cachep
->flags
& SLAB_RED_ZONE
) {
3007 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3008 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3009 slab_error(cachep
, "double free, or memory outside"
3010 " object was overwritten");
3012 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3013 objp
, *dbg_redzone1(cachep
, objp
),
3014 *dbg_redzone2(cachep
, objp
));
3016 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3017 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3019 #ifdef CONFIG_DEBUG_SLAB_LEAK
3024 slabp
= page_get_slab(virt_to_page(objp
));
3025 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3026 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3029 objp
+= obj_offset(cachep
);
3030 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
3031 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
3033 if (!(flags
& __GFP_WAIT
))
3034 ctor_flags
|= SLAB_CTOR_ATOMIC
;
3036 cachep
->ctor(objp
, cachep
, ctor_flags
);
3041 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3044 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3047 struct array_cache
*ac
;
3050 ac
= cpu_cache_get(cachep
);
3051 if (likely(ac
->avail
)) {
3052 STATS_INC_ALLOCHIT(cachep
);
3054 objp
= ac
->entry
[--ac
->avail
];
3056 STATS_INC_ALLOCMISS(cachep
);
3057 objp
= cache_alloc_refill(cachep
, flags
);
3062 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
3063 gfp_t flags
, void *caller
)
3065 unsigned long save_flags
;
3068 cache_alloc_debugcheck_before(cachep
, flags
);
3070 local_irq_save(save_flags
);
3072 if (unlikely(NUMA_BUILD
&&
3073 current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
)))
3074 objp
= alternate_node_alloc(cachep
, flags
);
3077 objp
= ____cache_alloc(cachep
, flags
);
3079 * We may just have run out of memory on the local node.
3080 * __cache_alloc_node() knows how to locate memory on other nodes
3082 if (NUMA_BUILD
&& !objp
)
3083 objp
= __cache_alloc_node(cachep
, flags
, numa_node_id());
3084 local_irq_restore(save_flags
);
3085 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
3093 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3095 * If we are in_interrupt, then process context, including cpusets and
3096 * mempolicy, may not apply and should not be used for allocation policy.
3098 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3100 int nid_alloc
, nid_here
;
3102 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3104 nid_alloc
= nid_here
= numa_node_id();
3105 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3106 nid_alloc
= cpuset_mem_spread_node();
3107 else if (current
->mempolicy
)
3108 nid_alloc
= slab_node(current
->mempolicy
);
3109 if (nid_alloc
!= nid_here
)
3110 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
3115 * Fallback function if there was no memory available and no objects on a
3116 * certain node and we are allowed to fall back. We mimick the behavior of
3117 * the page allocator. We fall back according to a zonelist determined by
3118 * the policy layer while obeying cpuset constraints.
3120 void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3122 struct zonelist
*zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
3123 ->node_zonelists
[gfp_zone(flags
)];
3127 for (z
= zonelist
->zones
; *z
&& !obj
; z
++)
3128 if (zone_idx(*z
) <= ZONE_NORMAL
&&
3129 cpuset_zone_allowed(*z
, flags
))
3130 obj
= __cache_alloc_node(cache
,
3131 flags
| __GFP_THISNODE
,
3137 * A interface to enable slab creation on nodeid
3139 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3142 struct list_head
*entry
;
3144 struct kmem_list3
*l3
;
3148 l3
= cachep
->nodelists
[nodeid
];
3153 spin_lock(&l3
->list_lock
);
3154 entry
= l3
->slabs_partial
.next
;
3155 if (entry
== &l3
->slabs_partial
) {
3156 l3
->free_touched
= 1;
3157 entry
= l3
->slabs_free
.next
;
3158 if (entry
== &l3
->slabs_free
)
3162 slabp
= list_entry(entry
, struct slab
, list
);
3163 check_spinlock_acquired_node(cachep
, nodeid
);
3164 check_slabp(cachep
, slabp
);
3166 STATS_INC_NODEALLOCS(cachep
);
3167 STATS_INC_ACTIVE(cachep
);
3168 STATS_SET_HIGH(cachep
);
3170 BUG_ON(slabp
->inuse
== cachep
->num
);
3172 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3173 check_slabp(cachep
, slabp
);
3175 /* move slabp to correct slabp list: */
3176 list_del(&slabp
->list
);
3178 if (slabp
->free
== BUFCTL_END
)
3179 list_add(&slabp
->list
, &l3
->slabs_full
);
3181 list_add(&slabp
->list
, &l3
->slabs_partial
);
3183 spin_unlock(&l3
->list_lock
);
3187 spin_unlock(&l3
->list_lock
);
3188 x
= cache_grow(cachep
, flags
, nodeid
);
3192 if (!(flags
& __GFP_THISNODE
))
3193 /* Unable to grow the cache. Fall back to other nodes. */
3194 return fallback_alloc(cachep
, flags
);
3204 * Caller needs to acquire correct kmem_list's list_lock
3206 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3210 struct kmem_list3
*l3
;
3212 for (i
= 0; i
< nr_objects
; i
++) {
3213 void *objp
= objpp
[i
];
3216 slabp
= virt_to_slab(objp
);
3217 l3
= cachep
->nodelists
[node
];
3218 list_del(&slabp
->list
);
3219 check_spinlock_acquired_node(cachep
, node
);
3220 check_slabp(cachep
, slabp
);
3221 slab_put_obj(cachep
, slabp
, objp
, node
);
3222 STATS_DEC_ACTIVE(cachep
);
3224 check_slabp(cachep
, slabp
);
3226 /* fixup slab chains */
3227 if (slabp
->inuse
== 0) {
3228 if (l3
->free_objects
> l3
->free_limit
) {
3229 l3
->free_objects
-= cachep
->num
;
3230 /* No need to drop any previously held
3231 * lock here, even if we have a off-slab slab
3232 * descriptor it is guaranteed to come from
3233 * a different cache, refer to comments before
3236 slab_destroy(cachep
, slabp
);
3238 list_add(&slabp
->list
, &l3
->slabs_free
);
3241 /* Unconditionally move a slab to the end of the
3242 * partial list on free - maximum time for the
3243 * other objects to be freed, too.
3245 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3250 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3253 struct kmem_list3
*l3
;
3254 int node
= numa_node_id();
3256 batchcount
= ac
->batchcount
;
3258 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3261 l3
= cachep
->nodelists
[node
];
3262 spin_lock(&l3
->list_lock
);
3264 struct array_cache
*shared_array
= l3
->shared
;
3265 int max
= shared_array
->limit
- shared_array
->avail
;
3267 if (batchcount
> max
)
3269 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3270 ac
->entry
, sizeof(void *) * batchcount
);
3271 shared_array
->avail
+= batchcount
;
3276 free_block(cachep
, ac
->entry
, batchcount
, node
);
3281 struct list_head
*p
;
3283 p
= l3
->slabs_free
.next
;
3284 while (p
!= &(l3
->slabs_free
)) {
3287 slabp
= list_entry(p
, struct slab
, list
);
3288 BUG_ON(slabp
->inuse
);
3293 STATS_SET_FREEABLE(cachep
, i
);
3296 spin_unlock(&l3
->list_lock
);
3297 ac
->avail
-= batchcount
;
3298 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3302 * Release an obj back to its cache. If the obj has a constructed state, it must
3303 * be in this state _before_ it is released. Called with disabled ints.
3305 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3307 struct array_cache
*ac
= cpu_cache_get(cachep
);
3310 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3312 if (cache_free_alien(cachep
, objp
))
3315 if (likely(ac
->avail
< ac
->limit
)) {
3316 STATS_INC_FREEHIT(cachep
);
3317 ac
->entry
[ac
->avail
++] = objp
;
3320 STATS_INC_FREEMISS(cachep
);
3321 cache_flusharray(cachep
, ac
);
3322 ac
->entry
[ac
->avail
++] = objp
;
3327 * kmem_cache_alloc - Allocate an object
3328 * @cachep: The cache to allocate from.
3329 * @flags: See kmalloc().
3331 * Allocate an object from this cache. The flags are only relevant
3332 * if the cache has no available objects.
3334 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3336 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3338 EXPORT_SYMBOL(kmem_cache_alloc
);
3341 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3342 * @cache: The cache to allocate from.
3343 * @flags: See kmalloc().
3345 * Allocate an object from this cache and set the allocated memory to zero.
3346 * The flags are only relevant if the cache has no available objects.
3348 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3350 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3352 memset(ret
, 0, obj_size(cache
));
3355 EXPORT_SYMBOL(kmem_cache_zalloc
);
3358 * kmem_ptr_validate - check if an untrusted pointer might
3360 * @cachep: the cache we're checking against
3361 * @ptr: pointer to validate
3363 * This verifies that the untrusted pointer looks sane:
3364 * it is _not_ a guarantee that the pointer is actually
3365 * part of the slab cache in question, but it at least
3366 * validates that the pointer can be dereferenced and
3367 * looks half-way sane.
3369 * Currently only used for dentry validation.
3371 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3373 unsigned long addr
= (unsigned long)ptr
;
3374 unsigned long min_addr
= PAGE_OFFSET
;
3375 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3376 unsigned long size
= cachep
->buffer_size
;
3379 if (unlikely(addr
< min_addr
))
3381 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3383 if (unlikely(addr
& align_mask
))
3385 if (unlikely(!kern_addr_valid(addr
)))
3387 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3389 page
= virt_to_page(ptr
);
3390 if (unlikely(!PageSlab(page
)))
3392 if (unlikely(page_get_cache(page
) != cachep
))
3401 * kmem_cache_alloc_node - Allocate an object on the specified node
3402 * @cachep: The cache to allocate from.
3403 * @flags: See kmalloc().
3404 * @nodeid: node number of the target node.
3406 * Identical to kmem_cache_alloc, except that this function is slow
3407 * and can sleep. And it will allocate memory on the given node, which
3408 * can improve the performance for cpu bound structures.
3409 * New and improved: it will now make sure that the object gets
3410 * put on the correct node list so that there is no false sharing.
3412 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3414 unsigned long save_flags
;
3417 cache_alloc_debugcheck_before(cachep
, flags
);
3418 local_irq_save(save_flags
);
3420 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3421 !cachep
->nodelists
[nodeid
])
3422 ptr
= ____cache_alloc(cachep
, flags
);
3424 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3425 local_irq_restore(save_flags
);
3427 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3428 __builtin_return_address(0));
3432 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3434 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3436 struct kmem_cache
*cachep
;
3438 cachep
= kmem_find_general_cachep(size
, flags
);
3439 if (unlikely(cachep
== NULL
))
3441 return kmem_cache_alloc_node(cachep
, flags
, node
);
3443 EXPORT_SYMBOL(__kmalloc_node
);
3447 * __do_kmalloc - allocate memory
3448 * @size: how many bytes of memory are required.
3449 * @flags: the type of memory to allocate (see kmalloc).
3450 * @caller: function caller for debug tracking of the caller
3452 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3455 struct kmem_cache
*cachep
;
3457 /* If you want to save a few bytes .text space: replace
3459 * Then kmalloc uses the uninlined functions instead of the inline
3462 cachep
= __find_general_cachep(size
, flags
);
3463 if (unlikely(cachep
== NULL
))
3465 return __cache_alloc(cachep
, flags
, caller
);
3469 void *__kmalloc(size_t size
, gfp_t flags
)
3471 #ifndef CONFIG_DEBUG_SLAB
3472 return __do_kmalloc(size
, flags
, NULL
);
3474 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3477 EXPORT_SYMBOL(__kmalloc
);
3479 #ifdef CONFIG_DEBUG_SLAB
3480 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3482 return __do_kmalloc(size
, flags
, caller
);
3484 EXPORT_SYMBOL(__kmalloc_track_caller
);
3488 * kmem_cache_free - Deallocate an object
3489 * @cachep: The cache the allocation was from.
3490 * @objp: The previously allocated object.
3492 * Free an object which was previously allocated from this
3495 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3497 unsigned long flags
;
3499 BUG_ON(virt_to_cache(objp
) != cachep
);
3501 local_irq_save(flags
);
3502 __cache_free(cachep
, objp
);
3503 local_irq_restore(flags
);
3505 EXPORT_SYMBOL(kmem_cache_free
);
3508 * kfree - free previously allocated memory
3509 * @objp: pointer returned by kmalloc.
3511 * If @objp is NULL, no operation is performed.
3513 * Don't free memory not originally allocated by kmalloc()
3514 * or you will run into trouble.
3516 void kfree(const void *objp
)
3518 struct kmem_cache
*c
;
3519 unsigned long flags
;
3521 if (unlikely(!objp
))
3523 local_irq_save(flags
);
3524 kfree_debugcheck(objp
);
3525 c
= virt_to_cache(objp
);
3526 debug_check_no_locks_freed(objp
, obj_size(c
));
3527 __cache_free(c
, (void *)objp
);
3528 local_irq_restore(flags
);
3530 EXPORT_SYMBOL(kfree
);
3532 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3534 return obj_size(cachep
);
3536 EXPORT_SYMBOL(kmem_cache_size
);
3538 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3540 return cachep
->name
;
3542 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3545 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3547 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3550 struct kmem_list3
*l3
;
3551 struct array_cache
*new_shared
;
3552 struct array_cache
**new_alien
;
3554 for_each_online_node(node
) {
3556 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3560 new_shared
= alloc_arraycache(node
,
3561 cachep
->shared
*cachep
->batchcount
,
3564 free_alien_cache(new_alien
);
3568 l3
= cachep
->nodelists
[node
];
3570 struct array_cache
*shared
= l3
->shared
;
3572 spin_lock_irq(&l3
->list_lock
);
3575 free_block(cachep
, shared
->entry
,
3576 shared
->avail
, node
);
3578 l3
->shared
= new_shared
;
3580 l3
->alien
= new_alien
;
3583 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3584 cachep
->batchcount
+ cachep
->num
;
3585 spin_unlock_irq(&l3
->list_lock
);
3587 free_alien_cache(new_alien
);
3590 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3592 free_alien_cache(new_alien
);
3597 kmem_list3_init(l3
);
3598 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3599 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3600 l3
->shared
= new_shared
;
3601 l3
->alien
= new_alien
;
3602 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3603 cachep
->batchcount
+ cachep
->num
;
3604 cachep
->nodelists
[node
] = l3
;
3609 if (!cachep
->next
.next
) {
3610 /* Cache is not active yet. Roll back what we did */
3613 if (cachep
->nodelists
[node
]) {
3614 l3
= cachep
->nodelists
[node
];
3617 free_alien_cache(l3
->alien
);
3619 cachep
->nodelists
[node
] = NULL
;
3627 struct ccupdate_struct
{
3628 struct kmem_cache
*cachep
;
3629 struct array_cache
*new[NR_CPUS
];
3632 static void do_ccupdate_local(void *info
)
3634 struct ccupdate_struct
*new = info
;
3635 struct array_cache
*old
;
3638 old
= cpu_cache_get(new->cachep
);
3640 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3641 new->new[smp_processor_id()] = old
;
3644 /* Always called with the cache_chain_mutex held */
3645 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3646 int batchcount
, int shared
)
3648 struct ccupdate_struct
*new;
3651 new = kzalloc(sizeof(*new), GFP_KERNEL
);
3655 for_each_online_cpu(i
) {
3656 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3659 for (i
--; i
>= 0; i
--)
3665 new->cachep
= cachep
;
3667 on_each_cpu(do_ccupdate_local
, (void *)new, 1, 1);
3670 cachep
->batchcount
= batchcount
;
3671 cachep
->limit
= limit
;
3672 cachep
->shared
= shared
;
3674 for_each_online_cpu(i
) {
3675 struct array_cache
*ccold
= new->new[i
];
3678 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3679 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3680 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3684 return alloc_kmemlist(cachep
);
3687 /* Called with cache_chain_mutex held always */
3688 static int enable_cpucache(struct kmem_cache
*cachep
)
3694 * The head array serves three purposes:
3695 * - create a LIFO ordering, i.e. return objects that are cache-warm
3696 * - reduce the number of spinlock operations.
3697 * - reduce the number of linked list operations on the slab and
3698 * bufctl chains: array operations are cheaper.
3699 * The numbers are guessed, we should auto-tune as described by
3702 if (cachep
->buffer_size
> 131072)
3704 else if (cachep
->buffer_size
> PAGE_SIZE
)
3706 else if (cachep
->buffer_size
> 1024)
3708 else if (cachep
->buffer_size
> 256)
3714 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3715 * allocation behaviour: Most allocs on one cpu, most free operations
3716 * on another cpu. For these cases, an efficient object passing between
3717 * cpus is necessary. This is provided by a shared array. The array
3718 * replaces Bonwick's magazine layer.
3719 * On uniprocessor, it's functionally equivalent (but less efficient)
3720 * to a larger limit. Thus disabled by default.
3724 if (cachep
->buffer_size
<= PAGE_SIZE
)
3730 * With debugging enabled, large batchcount lead to excessively long
3731 * periods with disabled local interrupts. Limit the batchcount
3736 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3738 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3739 cachep
->name
, -err
);
3744 * Drain an array if it contains any elements taking the l3 lock only if
3745 * necessary. Note that the l3 listlock also protects the array_cache
3746 * if drain_array() is used on the shared array.
3748 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3749 struct array_cache
*ac
, int force
, int node
)
3753 if (!ac
|| !ac
->avail
)
3755 if (ac
->touched
&& !force
) {
3758 spin_lock_irq(&l3
->list_lock
);
3760 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3761 if (tofree
> ac
->avail
)
3762 tofree
= (ac
->avail
+ 1) / 2;
3763 free_block(cachep
, ac
->entry
, tofree
, node
);
3764 ac
->avail
-= tofree
;
3765 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3766 sizeof(void *) * ac
->avail
);
3768 spin_unlock_irq(&l3
->list_lock
);
3773 * cache_reap - Reclaim memory from caches.
3774 * @unused: unused parameter
3776 * Called from workqueue/eventd every few seconds.
3778 * - clear the per-cpu caches for this CPU.
3779 * - return freeable pages to the main free memory pool.
3781 * If we cannot acquire the cache chain mutex then just give up - we'll try
3782 * again on the next iteration.
3784 static void cache_reap(void *unused
)
3786 struct kmem_cache
*searchp
;
3787 struct kmem_list3
*l3
;
3788 int node
= numa_node_id();
3790 if (!mutex_trylock(&cache_chain_mutex
)) {
3791 /* Give up. Setup the next iteration. */
3792 schedule_delayed_work(&__get_cpu_var(reap_work
),
3797 list_for_each_entry(searchp
, &cache_chain
, next
) {
3801 * We only take the l3 lock if absolutely necessary and we
3802 * have established with reasonable certainty that
3803 * we can do some work if the lock was obtained.
3805 l3
= searchp
->nodelists
[node
];
3807 reap_alien(searchp
, l3
);
3809 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3812 * These are racy checks but it does not matter
3813 * if we skip one check or scan twice.
3815 if (time_after(l3
->next_reap
, jiffies
))
3818 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3820 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3822 if (l3
->free_touched
)
3823 l3
->free_touched
= 0;
3827 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
3828 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3829 STATS_ADD_REAPED(searchp
, freed
);
3835 mutex_unlock(&cache_chain_mutex
);
3837 refresh_cpu_vm_stats(smp_processor_id());
3838 /* Set up the next iteration */
3839 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3842 #ifdef CONFIG_PROC_FS
3844 static void print_slabinfo_header(struct seq_file
*m
)
3847 * Output format version, so at least we can change it
3848 * without _too_ many complaints.
3851 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3853 seq_puts(m
, "slabinfo - version: 2.1\n");
3855 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3856 "<objperslab> <pagesperslab>");
3857 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3858 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3860 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3861 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3862 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3867 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3870 struct list_head
*p
;
3872 mutex_lock(&cache_chain_mutex
);
3874 print_slabinfo_header(m
);
3875 p
= cache_chain
.next
;
3878 if (p
== &cache_chain
)
3881 return list_entry(p
, struct kmem_cache
, next
);
3884 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3886 struct kmem_cache
*cachep
= p
;
3888 return cachep
->next
.next
== &cache_chain
?
3889 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3892 static void s_stop(struct seq_file
*m
, void *p
)
3894 mutex_unlock(&cache_chain_mutex
);
3897 static int s_show(struct seq_file
*m
, void *p
)
3899 struct kmem_cache
*cachep
= p
;
3901 unsigned long active_objs
;
3902 unsigned long num_objs
;
3903 unsigned long active_slabs
= 0;
3904 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3908 struct kmem_list3
*l3
;
3912 for_each_online_node(node
) {
3913 l3
= cachep
->nodelists
[node
];
3918 spin_lock_irq(&l3
->list_lock
);
3920 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
3921 if (slabp
->inuse
!= cachep
->num
&& !error
)
3922 error
= "slabs_full accounting error";
3923 active_objs
+= cachep
->num
;
3926 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
3927 if (slabp
->inuse
== cachep
->num
&& !error
)
3928 error
= "slabs_partial inuse accounting error";
3929 if (!slabp
->inuse
&& !error
)
3930 error
= "slabs_partial/inuse accounting error";
3931 active_objs
+= slabp
->inuse
;
3934 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
3935 if (slabp
->inuse
&& !error
)
3936 error
= "slabs_free/inuse accounting error";
3939 free_objects
+= l3
->free_objects
;
3941 shared_avail
+= l3
->shared
->avail
;
3943 spin_unlock_irq(&l3
->list_lock
);
3945 num_slabs
+= active_slabs
;
3946 num_objs
= num_slabs
* cachep
->num
;
3947 if (num_objs
- active_objs
!= free_objects
&& !error
)
3948 error
= "free_objects accounting error";
3950 name
= cachep
->name
;
3952 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3954 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3955 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3956 cachep
->num
, (1 << cachep
->gfporder
));
3957 seq_printf(m
, " : tunables %4u %4u %4u",
3958 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3959 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3960 active_slabs
, num_slabs
, shared_avail
);
3963 unsigned long high
= cachep
->high_mark
;
3964 unsigned long allocs
= cachep
->num_allocations
;
3965 unsigned long grown
= cachep
->grown
;
3966 unsigned long reaped
= cachep
->reaped
;
3967 unsigned long errors
= cachep
->errors
;
3968 unsigned long max_freeable
= cachep
->max_freeable
;
3969 unsigned long node_allocs
= cachep
->node_allocs
;
3970 unsigned long node_frees
= cachep
->node_frees
;
3971 unsigned long overflows
= cachep
->node_overflow
;
3973 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3974 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3975 reaped
, errors
, max_freeable
, node_allocs
,
3976 node_frees
, overflows
);
3980 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3981 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3982 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3983 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3985 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3986 allochit
, allocmiss
, freehit
, freemiss
);
3994 * slabinfo_op - iterator that generates /proc/slabinfo
4003 * num-pages-per-slab
4004 * + further values on SMP and with statistics enabled
4007 struct seq_operations slabinfo_op
= {
4014 #define MAX_SLABINFO_WRITE 128
4016 * slabinfo_write - Tuning for the slab allocator
4018 * @buffer: user buffer
4019 * @count: data length
4022 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4023 size_t count
, loff_t
*ppos
)
4025 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4026 int limit
, batchcount
, shared
, res
;
4027 struct kmem_cache
*cachep
;
4029 if (count
> MAX_SLABINFO_WRITE
)
4031 if (copy_from_user(&kbuf
, buffer
, count
))
4033 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4035 tmp
= strchr(kbuf
, ' ');
4040 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4043 /* Find the cache in the chain of caches. */
4044 mutex_lock(&cache_chain_mutex
);
4046 list_for_each_entry(cachep
, &cache_chain
, next
) {
4047 if (!strcmp(cachep
->name
, kbuf
)) {
4048 if (limit
< 1 || batchcount
< 1 ||
4049 batchcount
> limit
|| shared
< 0) {
4052 res
= do_tune_cpucache(cachep
, limit
,
4053 batchcount
, shared
);
4058 mutex_unlock(&cache_chain_mutex
);
4064 #ifdef CONFIG_DEBUG_SLAB_LEAK
4066 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4069 struct list_head
*p
;
4071 mutex_lock(&cache_chain_mutex
);
4072 p
= cache_chain
.next
;
4075 if (p
== &cache_chain
)
4078 return list_entry(p
, struct kmem_cache
, next
);
4081 static inline int add_caller(unsigned long *n
, unsigned long v
)
4091 unsigned long *q
= p
+ 2 * i
;
4105 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4111 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4117 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4118 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4120 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4125 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4127 #ifdef CONFIG_KALLSYMS
4130 unsigned long offset
, size
;
4131 char namebuf
[KSYM_NAME_LEN
+1];
4133 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4136 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4138 seq_printf(m
, " [%s]", modname
);
4142 seq_printf(m
, "%p", (void *)address
);
4145 static int leaks_show(struct seq_file
*m
, void *p
)
4147 struct kmem_cache
*cachep
= p
;
4149 struct kmem_list3
*l3
;
4151 unsigned long *n
= m
->private;
4155 if (!(cachep
->flags
& SLAB_STORE_USER
))
4157 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4160 /* OK, we can do it */
4164 for_each_online_node(node
) {
4165 l3
= cachep
->nodelists
[node
];
4170 spin_lock_irq(&l3
->list_lock
);
4172 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4173 handle_slab(n
, cachep
, slabp
);
4174 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4175 handle_slab(n
, cachep
, slabp
);
4176 spin_unlock_irq(&l3
->list_lock
);
4178 name
= cachep
->name
;
4180 /* Increase the buffer size */
4181 mutex_unlock(&cache_chain_mutex
);
4182 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4184 /* Too bad, we are really out */
4186 mutex_lock(&cache_chain_mutex
);
4189 *(unsigned long *)m
->private = n
[0] * 2;
4191 mutex_lock(&cache_chain_mutex
);
4192 /* Now make sure this entry will be retried */
4196 for (i
= 0; i
< n
[1]; i
++) {
4197 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4198 show_symbol(m
, n
[2*i
+2]);
4205 struct seq_operations slabstats_op
= {
4206 .start
= leaks_start
,
4215 * ksize - get the actual amount of memory allocated for a given object
4216 * @objp: Pointer to the object
4218 * kmalloc may internally round up allocations and return more memory
4219 * than requested. ksize() can be used to determine the actual amount of
4220 * memory allocated. The caller may use this additional memory, even though
4221 * a smaller amount of memory was initially specified with the kmalloc call.
4222 * The caller must guarantee that objp points to a valid object previously
4223 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4224 * must not be freed during the duration of the call.
4226 unsigned int ksize(const void *objp
)
4228 if (unlikely(objp
== NULL
))
4231 return obj_size(virt_to_cache(objp
));