3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
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/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
121 #include <asm/cacheflush.h>
122 #include <asm/tlbflush.h>
123 #include <asm/page.h>
125 #include <trace/events/kmem.h>
127 #include "internal.h"
130 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
131 * 0 for faster, smaller code (especially in the critical paths).
133 * STATS - 1 to collect stats for /proc/slabinfo.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
139 #ifdef CONFIG_DEBUG_SLAB
142 #define FORCED_DEBUG 1
146 #define FORCED_DEBUG 0
149 /* Shouldn't this be in a header file somewhere? */
150 #define BYTES_PER_WORD sizeof(void *)
151 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
153 #ifndef ARCH_KMALLOC_FLAGS
154 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158 * true if a page was allocated from pfmemalloc reserves for network-based
161 static bool pfmemalloc_active __read_mostly
;
163 /* Legal flag mask for kmem_cache_create(). */
165 # define CREATE_MASK (SLAB_RED_ZONE | \
166 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
169 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
170 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
171 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
173 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
176 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
177 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
183 * Bufctl's are used for linking objs within a slab
186 * This implementation relies on "struct page" for locating the cache &
187 * slab an object belongs to.
188 * This allows the bufctl structure to be small (one int), but limits
189 * the number of objects a slab (not a cache) can contain when off-slab
190 * bufctls are used. The limit is the size of the largest general cache
191 * that does not use off-slab slabs.
192 * For 32bit archs with 4 kB pages, is this 56.
193 * This is not serious, as it is only for large objects, when it is unwise
194 * to have too many per slab.
195 * Note: This limit can be raised by introducing a general cache whose size
196 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
199 typedef unsigned int kmem_bufctl_t
;
200 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
201 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
202 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
203 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
208 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
209 * arrange for kmem_freepages to be called via RCU. This is useful if
210 * we need to approach a kernel structure obliquely, from its address
211 * obtained without the usual locking. We can lock the structure to
212 * stabilize it and check it's still at the given address, only if we
213 * can be sure that the memory has not been meanwhile reused for some
214 * other kind of object (which our subsystem's lock might corrupt).
216 * rcu_read_lock before reading the address, then rcu_read_unlock after
217 * taking the spinlock within the structure expected at that address.
220 struct rcu_head head
;
221 struct kmem_cache
*cachep
;
228 * Manages the objs in a slab. Placed either at the beginning of mem allocated
229 * for a slab, or allocated from an general cache.
230 * Slabs are chained into three list: fully used, partial, fully free slabs.
235 struct list_head list
;
236 unsigned long colouroff
;
237 void *s_mem
; /* including colour offset */
238 unsigned int inuse
; /* num of objs active in slab */
240 unsigned short nodeid
;
242 struct slab_rcu __slab_cover_slab_rcu
;
250 * - LIFO ordering, to hand out cache-warm objects from _alloc
251 * - reduce the number of linked list operations
252 * - reduce spinlock operations
254 * The limit is stored in the per-cpu structure to reduce the data cache
261 unsigned int batchcount
;
262 unsigned int touched
;
265 * Must have this definition in here for the proper
266 * alignment of array_cache. Also simplifies accessing
269 * Entries should not be directly dereferenced as
270 * entries belonging to slabs marked pfmemalloc will
271 * have the lower bits set SLAB_OBJ_PFMEMALLOC
275 #define SLAB_OBJ_PFMEMALLOC 1
276 static inline bool is_obj_pfmemalloc(void *objp
)
278 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
281 static inline void set_obj_pfmemalloc(void **objp
)
283 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
287 static inline void clear_obj_pfmemalloc(void **objp
)
289 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
293 * bootstrap: The caches do not work without cpuarrays anymore, but the
294 * cpuarrays are allocated from the generic caches...
296 #define BOOT_CPUCACHE_ENTRIES 1
297 struct arraycache_init
{
298 struct array_cache cache
;
299 void *entries
[BOOT_CPUCACHE_ENTRIES
];
303 * The slab lists for all objects.
306 struct list_head slabs_partial
; /* partial list first, better asm code */
307 struct list_head slabs_full
;
308 struct list_head slabs_free
;
309 unsigned long free_objects
;
310 unsigned int free_limit
;
311 unsigned int colour_next
; /* Per-node cache coloring */
312 spinlock_t list_lock
;
313 struct array_cache
*shared
; /* shared per node */
314 struct array_cache
**alien
; /* on other nodes */
315 unsigned long next_reap
; /* updated without locking */
316 int free_touched
; /* updated without locking */
320 * Need this for bootstrapping a per node allocator.
322 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
323 static struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
324 #define CACHE_CACHE 0
325 #define SIZE_AC MAX_NUMNODES
326 #define SIZE_L3 (2 * MAX_NUMNODES)
328 static int drain_freelist(struct kmem_cache
*cache
,
329 struct kmem_list3
*l3
, int tofree
);
330 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
332 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
333 static void cache_reap(struct work_struct
*unused
);
336 * This function must be completely optimized away if a constant is passed to
337 * it. Mostly the same as what is in linux/slab.h except it returns an index.
339 static __always_inline
int index_of(const size_t size
)
341 extern void __bad_size(void);
343 if (__builtin_constant_p(size
)) {
351 #include <linux/kmalloc_sizes.h>
359 static int slab_early_init
= 1;
361 #define INDEX_AC index_of(sizeof(struct arraycache_init))
362 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
364 static void kmem_list3_init(struct kmem_list3
*parent
)
366 INIT_LIST_HEAD(&parent
->slabs_full
);
367 INIT_LIST_HEAD(&parent
->slabs_partial
);
368 INIT_LIST_HEAD(&parent
->slabs_free
);
369 parent
->shared
= NULL
;
370 parent
->alien
= NULL
;
371 parent
->colour_next
= 0;
372 spin_lock_init(&parent
->list_lock
);
373 parent
->free_objects
= 0;
374 parent
->free_touched
= 0;
377 #define MAKE_LIST(cachep, listp, slab, nodeid) \
379 INIT_LIST_HEAD(listp); \
380 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
383 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
385 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
386 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
387 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
390 #define CFLGS_OFF_SLAB (0x80000000UL)
391 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
393 #define BATCHREFILL_LIMIT 16
395 * Optimization question: fewer reaps means less probability for unnessary
396 * cpucache drain/refill cycles.
398 * OTOH the cpuarrays can contain lots of objects,
399 * which could lock up otherwise freeable slabs.
401 #define REAPTIMEOUT_CPUC (2*HZ)
402 #define REAPTIMEOUT_LIST3 (4*HZ)
405 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
406 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
407 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
408 #define STATS_INC_GROWN(x) ((x)->grown++)
409 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
410 #define STATS_SET_HIGH(x) \
412 if ((x)->num_active > (x)->high_mark) \
413 (x)->high_mark = (x)->num_active; \
415 #define STATS_INC_ERR(x) ((x)->errors++)
416 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
417 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
418 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
419 #define STATS_SET_FREEABLE(x, i) \
421 if ((x)->max_freeable < i) \
422 (x)->max_freeable = i; \
424 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
425 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
426 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
427 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
429 #define STATS_INC_ACTIVE(x) do { } while (0)
430 #define STATS_DEC_ACTIVE(x) do { } while (0)
431 #define STATS_INC_ALLOCED(x) do { } while (0)
432 #define STATS_INC_GROWN(x) do { } while (0)
433 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
434 #define STATS_SET_HIGH(x) do { } while (0)
435 #define STATS_INC_ERR(x) do { } while (0)
436 #define STATS_INC_NODEALLOCS(x) do { } while (0)
437 #define STATS_INC_NODEFREES(x) do { } while (0)
438 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
439 #define STATS_SET_FREEABLE(x, i) do { } while (0)
440 #define STATS_INC_ALLOCHIT(x) do { } while (0)
441 #define STATS_INC_ALLOCMISS(x) do { } while (0)
442 #define STATS_INC_FREEHIT(x) do { } while (0)
443 #define STATS_INC_FREEMISS(x) do { } while (0)
449 * memory layout of objects:
451 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
452 * the end of an object is aligned with the end of the real
453 * allocation. Catches writes behind the end of the allocation.
454 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
456 * cachep->obj_offset: The real object.
457 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
458 * cachep->size - 1* BYTES_PER_WORD: last caller address
459 * [BYTES_PER_WORD long]
461 static int obj_offset(struct kmem_cache
*cachep
)
463 return cachep
->obj_offset
;
466 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
468 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
469 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
470 sizeof(unsigned long long));
473 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
475 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
476 if (cachep
->flags
& SLAB_STORE_USER
)
477 return (unsigned long long *)(objp
+ cachep
->size
-
478 sizeof(unsigned long long) -
480 return (unsigned long long *) (objp
+ cachep
->size
-
481 sizeof(unsigned long long));
484 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
486 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
487 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
492 #define obj_offset(x) 0
493 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
494 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
495 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
499 #ifdef CONFIG_TRACING
500 size_t slab_buffer_size(struct kmem_cache
*cachep
)
504 EXPORT_SYMBOL(slab_buffer_size
);
508 * Do not go above this order unless 0 objects fit into the slab or
509 * overridden on the command line.
511 #define SLAB_MAX_ORDER_HI 1
512 #define SLAB_MAX_ORDER_LO 0
513 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
514 static bool slab_max_order_set __initdata
;
516 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
518 page
= compound_head(page
);
519 BUG_ON(!PageSlab(page
));
520 return page
->slab_cache
;
523 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
525 struct page
*page
= virt_to_head_page(obj
);
526 return page
->slab_cache
;
529 static inline struct slab
*virt_to_slab(const void *obj
)
531 struct page
*page
= virt_to_head_page(obj
);
533 VM_BUG_ON(!PageSlab(page
));
534 return page
->slab_page
;
537 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
540 return slab
->s_mem
+ cache
->size
* idx
;
544 * We want to avoid an expensive divide : (offset / cache->size)
545 * Using the fact that size is a constant for a particular cache,
546 * we can replace (offset / cache->size) by
547 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
549 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
550 const struct slab
*slab
, void *obj
)
552 u32 offset
= (obj
- slab
->s_mem
);
553 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
557 * These are the default caches for kmalloc. Custom caches can have other sizes.
559 struct cache_sizes malloc_sizes
[] = {
560 #define CACHE(x) { .cs_size = (x) },
561 #include <linux/kmalloc_sizes.h>
565 EXPORT_SYMBOL(malloc_sizes
);
567 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
573 static struct cache_names __initdata cache_names
[] = {
574 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
575 #include <linux/kmalloc_sizes.h>
580 static struct arraycache_init initarray_cache __initdata
=
581 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
582 static struct arraycache_init initarray_generic
=
583 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
585 /* internal cache of cache description objs */
586 static struct kmem_list3
*cache_cache_nodelists
[MAX_NUMNODES
];
587 static struct kmem_cache cache_cache
= {
588 .nodelists
= cache_cache_nodelists
,
590 .limit
= BOOT_CPUCACHE_ENTRIES
,
592 .size
= sizeof(struct kmem_cache
),
593 .name
= "kmem_cache",
596 #define BAD_ALIEN_MAGIC 0x01020304ul
598 #ifdef CONFIG_LOCKDEP
601 * Slab sometimes uses the kmalloc slabs to store the slab headers
602 * for other slabs "off slab".
603 * The locking for this is tricky in that it nests within the locks
604 * of all other slabs in a few places; to deal with this special
605 * locking we put on-slab caches into a separate lock-class.
607 * We set lock class for alien array caches which are up during init.
608 * The lock annotation will be lost if all cpus of a node goes down and
609 * then comes back up during hotplug
611 static struct lock_class_key on_slab_l3_key
;
612 static struct lock_class_key on_slab_alc_key
;
614 static struct lock_class_key debugobj_l3_key
;
615 static struct lock_class_key debugobj_alc_key
;
617 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
618 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
621 struct array_cache
**alc
;
622 struct kmem_list3
*l3
;
625 l3
= cachep
->nodelists
[q
];
629 lockdep_set_class(&l3
->list_lock
, l3_key
);
632 * FIXME: This check for BAD_ALIEN_MAGIC
633 * should go away when common slab code is taught to
634 * work even without alien caches.
635 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
636 * for alloc_alien_cache,
638 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
642 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
646 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
648 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
651 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
655 for_each_online_node(node
)
656 slab_set_debugobj_lock_classes_node(cachep
, node
);
659 static void init_node_lock_keys(int q
)
661 struct cache_sizes
*s
= malloc_sizes
;
666 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
667 struct kmem_list3
*l3
;
669 l3
= s
->cs_cachep
->nodelists
[q
];
670 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
673 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
674 &on_slab_alc_key
, q
);
678 static inline void init_lock_keys(void)
683 init_node_lock_keys(node
);
686 static void init_node_lock_keys(int q
)
690 static inline void init_lock_keys(void)
694 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
698 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
703 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
705 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
707 return cachep
->array
[smp_processor_id()];
710 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
713 struct cache_sizes
*csizep
= malloc_sizes
;
716 /* This happens if someone tries to call
717 * kmem_cache_create(), or __kmalloc(), before
718 * the generic caches are initialized.
720 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
723 return ZERO_SIZE_PTR
;
725 while (size
> csizep
->cs_size
)
729 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
730 * has cs_{dma,}cachep==NULL. Thus no special case
731 * for large kmalloc calls required.
733 #ifdef CONFIG_ZONE_DMA
734 if (unlikely(gfpflags
& GFP_DMA
))
735 return csizep
->cs_dmacachep
;
737 return csizep
->cs_cachep
;
740 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
742 return __find_general_cachep(size
, gfpflags
);
745 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
747 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
751 * Calculate the number of objects and left-over bytes for a given buffer size.
753 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
754 size_t align
, int flags
, size_t *left_over
,
759 size_t slab_size
= PAGE_SIZE
<< gfporder
;
762 * The slab management structure can be either off the slab or
763 * on it. For the latter case, the memory allocated for a
767 * - One kmem_bufctl_t for each object
768 * - Padding to respect alignment of @align
769 * - @buffer_size bytes for each object
771 * If the slab management structure is off the slab, then the
772 * alignment will already be calculated into the size. Because
773 * the slabs are all pages aligned, the objects will be at the
774 * correct alignment when allocated.
776 if (flags
& CFLGS_OFF_SLAB
) {
778 nr_objs
= slab_size
/ buffer_size
;
780 if (nr_objs
> SLAB_LIMIT
)
781 nr_objs
= SLAB_LIMIT
;
784 * Ignore padding for the initial guess. The padding
785 * is at most @align-1 bytes, and @buffer_size is at
786 * least @align. In the worst case, this result will
787 * be one greater than the number of objects that fit
788 * into the memory allocation when taking the padding
791 nr_objs
= (slab_size
- sizeof(struct slab
)) /
792 (buffer_size
+ sizeof(kmem_bufctl_t
));
795 * This calculated number will be either the right
796 * amount, or one greater than what we want.
798 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
802 if (nr_objs
> SLAB_LIMIT
)
803 nr_objs
= SLAB_LIMIT
;
805 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
808 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
811 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
813 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
816 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
817 function
, cachep
->name
, msg
);
822 * By default on NUMA we use alien caches to stage the freeing of
823 * objects allocated from other nodes. This causes massive memory
824 * inefficiencies when using fake NUMA setup to split memory into a
825 * large number of small nodes, so it can be disabled on the command
829 static int use_alien_caches __read_mostly
= 1;
830 static int __init
noaliencache_setup(char *s
)
832 use_alien_caches
= 0;
835 __setup("noaliencache", noaliencache_setup
);
837 static int __init
slab_max_order_setup(char *str
)
839 get_option(&str
, &slab_max_order
);
840 slab_max_order
= slab_max_order
< 0 ? 0 :
841 min(slab_max_order
, MAX_ORDER
- 1);
842 slab_max_order_set
= true;
846 __setup("slab_max_order=", slab_max_order_setup
);
850 * Special reaping functions for NUMA systems called from cache_reap().
851 * These take care of doing round robin flushing of alien caches (containing
852 * objects freed on different nodes from which they were allocated) and the
853 * flushing of remote pcps by calling drain_node_pages.
855 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
857 static void init_reap_node(int cpu
)
861 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
862 if (node
== MAX_NUMNODES
)
863 node
= first_node(node_online_map
);
865 per_cpu(slab_reap_node
, cpu
) = node
;
868 static void next_reap_node(void)
870 int node
= __this_cpu_read(slab_reap_node
);
872 node
= next_node(node
, node_online_map
);
873 if (unlikely(node
>= MAX_NUMNODES
))
874 node
= first_node(node_online_map
);
875 __this_cpu_write(slab_reap_node
, node
);
879 #define init_reap_node(cpu) do { } while (0)
880 #define next_reap_node(void) do { } while (0)
884 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
885 * via the workqueue/eventd.
886 * Add the CPU number into the expiration time to minimize the possibility of
887 * the CPUs getting into lockstep and contending for the global cache chain
890 static void __cpuinit
start_cpu_timer(int cpu
)
892 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
895 * When this gets called from do_initcalls via cpucache_init(),
896 * init_workqueues() has already run, so keventd will be setup
899 if (keventd_up() && reap_work
->work
.func
== NULL
) {
901 INIT_DELAYED_WORK_DEFERRABLE(reap_work
, cache_reap
);
902 schedule_delayed_work_on(cpu
, reap_work
,
903 __round_jiffies_relative(HZ
, cpu
));
907 static struct array_cache
*alloc_arraycache(int node
, int entries
,
908 int batchcount
, gfp_t gfp
)
910 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
911 struct array_cache
*nc
= NULL
;
913 nc
= kmalloc_node(memsize
, gfp
, node
);
915 * The array_cache structures contain pointers to free object.
916 * However, when such objects are allocated or transferred to another
917 * cache the pointers are not cleared and they could be counted as
918 * valid references during a kmemleak scan. Therefore, kmemleak must
919 * not scan such objects.
921 kmemleak_no_scan(nc
);
925 nc
->batchcount
= batchcount
;
927 spin_lock_init(&nc
->lock
);
932 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
934 struct page
*page
= virt_to_page(slabp
->s_mem
);
936 return PageSlabPfmemalloc(page
);
939 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
940 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
941 struct array_cache
*ac
)
943 struct kmem_list3
*l3
= cachep
->nodelists
[numa_mem_id()];
947 if (!pfmemalloc_active
)
950 spin_lock_irqsave(&l3
->list_lock
, flags
);
951 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
952 if (is_slab_pfmemalloc(slabp
))
955 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
956 if (is_slab_pfmemalloc(slabp
))
959 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
960 if (is_slab_pfmemalloc(slabp
))
963 pfmemalloc_active
= false;
965 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
968 static void *ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
969 gfp_t flags
, bool force_refill
)
972 void *objp
= ac
->entry
[--ac
->avail
];
974 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
975 if (unlikely(is_obj_pfmemalloc(objp
))) {
976 struct kmem_list3
*l3
;
978 if (gfp_pfmemalloc_allowed(flags
)) {
979 clear_obj_pfmemalloc(&objp
);
983 /* The caller cannot use PFMEMALLOC objects, find another one */
984 for (i
= 1; i
< ac
->avail
; i
++) {
985 /* If a !PFMEMALLOC object is found, swap them */
986 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
988 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
989 ac
->entry
[ac
->avail
] = objp
;
995 * If there are empty slabs on the slabs_free list and we are
996 * being forced to refill the cache, mark this one !pfmemalloc.
998 l3
= cachep
->nodelists
[numa_mem_id()];
999 if (!list_empty(&l3
->slabs_free
) && force_refill
) {
1000 struct slab
*slabp
= virt_to_slab(objp
);
1001 ClearPageSlabPfmemalloc(virt_to_page(slabp
->s_mem
));
1002 clear_obj_pfmemalloc(&objp
);
1003 recheck_pfmemalloc_active(cachep
, ac
);
1007 /* No !PFMEMALLOC objects available */
1015 static void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1018 if (unlikely(pfmemalloc_active
)) {
1019 /* Some pfmemalloc slabs exist, check if this is one */
1020 struct page
*page
= virt_to_page(objp
);
1021 if (PageSlabPfmemalloc(page
))
1022 set_obj_pfmemalloc(&objp
);
1025 ac
->entry
[ac
->avail
++] = objp
;
1029 * Transfer objects in one arraycache to another.
1030 * Locking must be handled by the caller.
1032 * Return the number of entries transferred.
1034 static int transfer_objects(struct array_cache
*to
,
1035 struct array_cache
*from
, unsigned int max
)
1037 /* Figure out how many entries to transfer */
1038 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
1043 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1044 sizeof(void *) *nr
);
1053 #define drain_alien_cache(cachep, alien) do { } while (0)
1054 #define reap_alien(cachep, l3) do { } while (0)
1056 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1058 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1061 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1065 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1070 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1076 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1077 gfp_t flags
, int nodeid
)
1082 #else /* CONFIG_NUMA */
1084 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1085 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1087 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1089 struct array_cache
**ac_ptr
;
1090 int memsize
= sizeof(void *) * nr_node_ids
;
1095 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1098 if (i
== node
|| !node_online(i
))
1100 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1102 for (i
--; i
>= 0; i
--)
1112 static void free_alien_cache(struct array_cache
**ac_ptr
)
1123 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1124 struct array_cache
*ac
, int node
)
1126 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1129 spin_lock(&rl3
->list_lock
);
1131 * Stuff objects into the remote nodes shared array first.
1132 * That way we could avoid the overhead of putting the objects
1133 * into the free lists and getting them back later.
1136 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1138 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1140 spin_unlock(&rl3
->list_lock
);
1145 * Called from cache_reap() to regularly drain alien caches round robin.
1147 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1149 int node
= __this_cpu_read(slab_reap_node
);
1152 struct array_cache
*ac
= l3
->alien
[node
];
1154 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1155 __drain_alien_cache(cachep
, ac
, node
);
1156 spin_unlock_irq(&ac
->lock
);
1161 static void drain_alien_cache(struct kmem_cache
*cachep
,
1162 struct array_cache
**alien
)
1165 struct array_cache
*ac
;
1166 unsigned long flags
;
1168 for_each_online_node(i
) {
1171 spin_lock_irqsave(&ac
->lock
, flags
);
1172 __drain_alien_cache(cachep
, ac
, i
);
1173 spin_unlock_irqrestore(&ac
->lock
, flags
);
1178 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1180 struct slab
*slabp
= virt_to_slab(objp
);
1181 int nodeid
= slabp
->nodeid
;
1182 struct kmem_list3
*l3
;
1183 struct array_cache
*alien
= NULL
;
1186 node
= numa_mem_id();
1189 * Make sure we are not freeing a object from another node to the array
1190 * cache on this cpu.
1192 if (likely(slabp
->nodeid
== node
))
1195 l3
= cachep
->nodelists
[node
];
1196 STATS_INC_NODEFREES(cachep
);
1197 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1198 alien
= l3
->alien
[nodeid
];
1199 spin_lock(&alien
->lock
);
1200 if (unlikely(alien
->avail
== alien
->limit
)) {
1201 STATS_INC_ACOVERFLOW(cachep
);
1202 __drain_alien_cache(cachep
, alien
, nodeid
);
1204 ac_put_obj(cachep
, alien
, objp
);
1205 spin_unlock(&alien
->lock
);
1207 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1208 free_block(cachep
, &objp
, 1, nodeid
);
1209 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1216 * Allocates and initializes nodelists for a node on each slab cache, used for
1217 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1218 * will be allocated off-node since memory is not yet online for the new node.
1219 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1222 * Must hold slab_mutex.
1224 static int init_cache_nodelists_node(int node
)
1226 struct kmem_cache
*cachep
;
1227 struct kmem_list3
*l3
;
1228 const int memsize
= sizeof(struct kmem_list3
);
1230 list_for_each_entry(cachep
, &slab_caches
, list
) {
1232 * Set up the size64 kmemlist for cpu before we can
1233 * begin anything. Make sure some other cpu on this
1234 * node has not already allocated this
1236 if (!cachep
->nodelists
[node
]) {
1237 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1240 kmem_list3_init(l3
);
1241 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1242 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1245 * The l3s don't come and go as CPUs come and
1246 * go. slab_mutex is sufficient
1249 cachep
->nodelists
[node
] = l3
;
1252 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1253 cachep
->nodelists
[node
]->free_limit
=
1254 (1 + nr_cpus_node(node
)) *
1255 cachep
->batchcount
+ cachep
->num
;
1256 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1261 static void __cpuinit
cpuup_canceled(long cpu
)
1263 struct kmem_cache
*cachep
;
1264 struct kmem_list3
*l3
= NULL
;
1265 int node
= cpu_to_mem(cpu
);
1266 const struct cpumask
*mask
= cpumask_of_node(node
);
1268 list_for_each_entry(cachep
, &slab_caches
, list
) {
1269 struct array_cache
*nc
;
1270 struct array_cache
*shared
;
1271 struct array_cache
**alien
;
1273 /* cpu is dead; no one can alloc from it. */
1274 nc
= cachep
->array
[cpu
];
1275 cachep
->array
[cpu
] = NULL
;
1276 l3
= cachep
->nodelists
[node
];
1279 goto free_array_cache
;
1281 spin_lock_irq(&l3
->list_lock
);
1283 /* Free limit for this kmem_list3 */
1284 l3
->free_limit
-= cachep
->batchcount
;
1286 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1288 if (!cpumask_empty(mask
)) {
1289 spin_unlock_irq(&l3
->list_lock
);
1290 goto free_array_cache
;
1293 shared
= l3
->shared
;
1295 free_block(cachep
, shared
->entry
,
1296 shared
->avail
, node
);
1303 spin_unlock_irq(&l3
->list_lock
);
1307 drain_alien_cache(cachep
, alien
);
1308 free_alien_cache(alien
);
1314 * In the previous loop, all the objects were freed to
1315 * the respective cache's slabs, now we can go ahead and
1316 * shrink each nodelist to its limit.
1318 list_for_each_entry(cachep
, &slab_caches
, list
) {
1319 l3
= cachep
->nodelists
[node
];
1322 drain_freelist(cachep
, l3
, l3
->free_objects
);
1326 static int __cpuinit
cpuup_prepare(long cpu
)
1328 struct kmem_cache
*cachep
;
1329 struct kmem_list3
*l3
= NULL
;
1330 int node
= cpu_to_mem(cpu
);
1334 * We need to do this right in the beginning since
1335 * alloc_arraycache's are going to use this list.
1336 * kmalloc_node allows us to add the slab to the right
1337 * kmem_list3 and not this cpu's kmem_list3
1339 err
= init_cache_nodelists_node(node
);
1344 * Now we can go ahead with allocating the shared arrays and
1347 list_for_each_entry(cachep
, &slab_caches
, list
) {
1348 struct array_cache
*nc
;
1349 struct array_cache
*shared
= NULL
;
1350 struct array_cache
**alien
= NULL
;
1352 nc
= alloc_arraycache(node
, cachep
->limit
,
1353 cachep
->batchcount
, GFP_KERNEL
);
1356 if (cachep
->shared
) {
1357 shared
= alloc_arraycache(node
,
1358 cachep
->shared
* cachep
->batchcount
,
1359 0xbaadf00d, GFP_KERNEL
);
1365 if (use_alien_caches
) {
1366 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1373 cachep
->array
[cpu
] = nc
;
1374 l3
= cachep
->nodelists
[node
];
1377 spin_lock_irq(&l3
->list_lock
);
1380 * We are serialised from CPU_DEAD or
1381 * CPU_UP_CANCELLED by the cpucontrol lock
1383 l3
->shared
= shared
;
1392 spin_unlock_irq(&l3
->list_lock
);
1394 free_alien_cache(alien
);
1395 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1396 slab_set_debugobj_lock_classes_node(cachep
, node
);
1398 init_node_lock_keys(node
);
1402 cpuup_canceled(cpu
);
1406 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1407 unsigned long action
, void *hcpu
)
1409 long cpu
= (long)hcpu
;
1413 case CPU_UP_PREPARE
:
1414 case CPU_UP_PREPARE_FROZEN
:
1415 mutex_lock(&slab_mutex
);
1416 err
= cpuup_prepare(cpu
);
1417 mutex_unlock(&slab_mutex
);
1420 case CPU_ONLINE_FROZEN
:
1421 start_cpu_timer(cpu
);
1423 #ifdef CONFIG_HOTPLUG_CPU
1424 case CPU_DOWN_PREPARE
:
1425 case CPU_DOWN_PREPARE_FROZEN
:
1427 * Shutdown cache reaper. Note that the slab_mutex is
1428 * held so that if cache_reap() is invoked it cannot do
1429 * anything expensive but will only modify reap_work
1430 * and reschedule the timer.
1432 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1433 /* Now the cache_reaper is guaranteed to be not running. */
1434 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1436 case CPU_DOWN_FAILED
:
1437 case CPU_DOWN_FAILED_FROZEN
:
1438 start_cpu_timer(cpu
);
1441 case CPU_DEAD_FROZEN
:
1443 * Even if all the cpus of a node are down, we don't free the
1444 * kmem_list3 of any cache. This to avoid a race between
1445 * cpu_down, and a kmalloc allocation from another cpu for
1446 * memory from the node of the cpu going down. The list3
1447 * structure is usually allocated from kmem_cache_create() and
1448 * gets destroyed at kmem_cache_destroy().
1452 case CPU_UP_CANCELED
:
1453 case CPU_UP_CANCELED_FROZEN
:
1454 mutex_lock(&slab_mutex
);
1455 cpuup_canceled(cpu
);
1456 mutex_unlock(&slab_mutex
);
1459 return notifier_from_errno(err
);
1462 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1463 &cpuup_callback
, NULL
, 0
1466 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1468 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1469 * Returns -EBUSY if all objects cannot be drained so that the node is not
1472 * Must hold slab_mutex.
1474 static int __meminit
drain_cache_nodelists_node(int node
)
1476 struct kmem_cache
*cachep
;
1479 list_for_each_entry(cachep
, &slab_caches
, list
) {
1480 struct kmem_list3
*l3
;
1482 l3
= cachep
->nodelists
[node
];
1486 drain_freelist(cachep
, l3
, l3
->free_objects
);
1488 if (!list_empty(&l3
->slabs_full
) ||
1489 !list_empty(&l3
->slabs_partial
)) {
1497 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1498 unsigned long action
, void *arg
)
1500 struct memory_notify
*mnb
= arg
;
1504 nid
= mnb
->status_change_nid
;
1509 case MEM_GOING_ONLINE
:
1510 mutex_lock(&slab_mutex
);
1511 ret
= init_cache_nodelists_node(nid
);
1512 mutex_unlock(&slab_mutex
);
1514 case MEM_GOING_OFFLINE
:
1515 mutex_lock(&slab_mutex
);
1516 ret
= drain_cache_nodelists_node(nid
);
1517 mutex_unlock(&slab_mutex
);
1521 case MEM_CANCEL_ONLINE
:
1522 case MEM_CANCEL_OFFLINE
:
1526 return notifier_from_errno(ret
);
1528 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1531 * swap the static kmem_list3 with kmalloced memory
1533 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1536 struct kmem_list3
*ptr
;
1538 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1541 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1543 * Do not assume that spinlocks can be initialized via memcpy:
1545 spin_lock_init(&ptr
->list_lock
);
1547 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1548 cachep
->nodelists
[nodeid
] = ptr
;
1552 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1553 * size of kmem_list3.
1555 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1559 for_each_online_node(node
) {
1560 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1561 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1563 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1568 * Initialisation. Called after the page allocator have been initialised and
1569 * before smp_init().
1571 void __init
kmem_cache_init(void)
1574 struct cache_sizes
*sizes
;
1575 struct cache_names
*names
;
1580 if (num_possible_nodes() == 1)
1581 use_alien_caches
= 0;
1583 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1584 kmem_list3_init(&initkmem_list3
[i
]);
1585 if (i
< MAX_NUMNODES
)
1586 cache_cache
.nodelists
[i
] = NULL
;
1588 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1591 * Fragmentation resistance on low memory - only use bigger
1592 * page orders on machines with more than 32MB of memory if
1593 * not overridden on the command line.
1595 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1596 slab_max_order
= SLAB_MAX_ORDER_HI
;
1598 /* Bootstrap is tricky, because several objects are allocated
1599 * from caches that do not exist yet:
1600 * 1) initialize the cache_cache cache: it contains the struct
1601 * kmem_cache structures of all caches, except cache_cache itself:
1602 * cache_cache is statically allocated.
1603 * Initially an __init data area is used for the head array and the
1604 * kmem_list3 structures, it's replaced with a kmalloc allocated
1605 * array at the end of the bootstrap.
1606 * 2) Create the first kmalloc cache.
1607 * The struct kmem_cache for the new cache is allocated normally.
1608 * An __init data area is used for the head array.
1609 * 3) Create the remaining kmalloc caches, with minimally sized
1611 * 4) Replace the __init data head arrays for cache_cache and the first
1612 * kmalloc cache with kmalloc allocated arrays.
1613 * 5) Replace the __init data for kmem_list3 for cache_cache and
1614 * the other cache's with kmalloc allocated memory.
1615 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1618 node
= numa_mem_id();
1620 /* 1) create the cache_cache */
1621 INIT_LIST_HEAD(&slab_caches
);
1622 list_add(&cache_cache
.list
, &slab_caches
);
1623 cache_cache
.colour_off
= cache_line_size();
1624 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1625 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1628 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1630 cache_cache
.size
= offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1631 nr_node_ids
* sizeof(struct kmem_list3
*);
1632 cache_cache
.object_size
= cache_cache
.size
;
1633 cache_cache
.size
= ALIGN(cache_cache
.size
,
1635 cache_cache
.reciprocal_buffer_size
=
1636 reciprocal_value(cache_cache
.size
);
1638 for (order
= 0; order
< MAX_ORDER
; order
++) {
1639 cache_estimate(order
, cache_cache
.size
,
1640 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1641 if (cache_cache
.num
)
1644 BUG_ON(!cache_cache
.num
);
1645 cache_cache
.gfporder
= order
;
1646 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1647 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1648 sizeof(struct slab
), cache_line_size());
1650 /* 2+3) create the kmalloc caches */
1651 sizes
= malloc_sizes
;
1652 names
= cache_names
;
1655 * Initialize the caches that provide memory for the array cache and the
1656 * kmem_list3 structures first. Without this, further allocations will
1660 sizes
[INDEX_AC
].cs_cachep
= __kmem_cache_create(names
[INDEX_AC
].name
,
1661 sizes
[INDEX_AC
].cs_size
,
1662 ARCH_KMALLOC_MINALIGN
,
1663 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1666 if (INDEX_AC
!= INDEX_L3
) {
1667 sizes
[INDEX_L3
].cs_cachep
=
1668 __kmem_cache_create(names
[INDEX_L3
].name
,
1669 sizes
[INDEX_L3
].cs_size
,
1670 ARCH_KMALLOC_MINALIGN
,
1671 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1675 slab_early_init
= 0;
1677 while (sizes
->cs_size
!= ULONG_MAX
) {
1679 * For performance, all the general caches are L1 aligned.
1680 * This should be particularly beneficial on SMP boxes, as it
1681 * eliminates "false sharing".
1682 * Note for systems short on memory removing the alignment will
1683 * allow tighter packing of the smaller caches.
1685 if (!sizes
->cs_cachep
) {
1686 sizes
->cs_cachep
= __kmem_cache_create(names
->name
,
1688 ARCH_KMALLOC_MINALIGN
,
1689 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1692 #ifdef CONFIG_ZONE_DMA
1693 sizes
->cs_dmacachep
= __kmem_cache_create(
1696 ARCH_KMALLOC_MINALIGN
,
1697 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1704 /* 4) Replace the bootstrap head arrays */
1706 struct array_cache
*ptr
;
1708 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1710 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1711 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1712 sizeof(struct arraycache_init
));
1714 * Do not assume that spinlocks can be initialized via memcpy:
1716 spin_lock_init(&ptr
->lock
);
1718 cache_cache
.array
[smp_processor_id()] = ptr
;
1720 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1722 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1723 != &initarray_generic
.cache
);
1724 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1725 sizeof(struct arraycache_init
));
1727 * Do not assume that spinlocks can be initialized via memcpy:
1729 spin_lock_init(&ptr
->lock
);
1731 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1734 /* 5) Replace the bootstrap kmem_list3's */
1738 for_each_online_node(nid
) {
1739 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1741 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1742 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1744 if (INDEX_AC
!= INDEX_L3
) {
1745 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1746 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1754 void __init
kmem_cache_init_late(void)
1756 struct kmem_cache
*cachep
;
1760 /* Annotate slab for lockdep -- annotate the malloc caches */
1763 /* 6) resize the head arrays to their final sizes */
1764 mutex_lock(&slab_mutex
);
1765 list_for_each_entry(cachep
, &slab_caches
, list
)
1766 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1768 mutex_unlock(&slab_mutex
);
1774 * Register a cpu startup notifier callback that initializes
1775 * cpu_cache_get for all new cpus
1777 register_cpu_notifier(&cpucache_notifier
);
1781 * Register a memory hotplug callback that initializes and frees
1784 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1788 * The reap timers are started later, with a module init call: That part
1789 * of the kernel is not yet operational.
1793 static int __init
cpucache_init(void)
1798 * Register the timers that return unneeded pages to the page allocator
1800 for_each_online_cpu(cpu
)
1801 start_cpu_timer(cpu
);
1807 __initcall(cpucache_init
);
1809 static noinline
void
1810 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1812 struct kmem_list3
*l3
;
1814 unsigned long flags
;
1818 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1820 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1821 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1823 for_each_online_node(node
) {
1824 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1825 unsigned long active_slabs
= 0, num_slabs
= 0;
1827 l3
= cachep
->nodelists
[node
];
1831 spin_lock_irqsave(&l3
->list_lock
, flags
);
1832 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
1833 active_objs
+= cachep
->num
;
1836 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
1837 active_objs
+= slabp
->inuse
;
1840 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
1843 free_objects
+= l3
->free_objects
;
1844 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
1846 num_slabs
+= active_slabs
;
1847 num_objs
= num_slabs
* cachep
->num
;
1849 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1850 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1856 * Interface to system's page allocator. No need to hold the cache-lock.
1858 * If we requested dmaable memory, we will get it. Even if we
1859 * did not request dmaable memory, we might get it, but that
1860 * would be relatively rare and ignorable.
1862 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1870 * Nommu uses slab's for process anonymous memory allocations, and thus
1871 * requires __GFP_COMP to properly refcount higher order allocations
1873 flags
|= __GFP_COMP
;
1876 flags
|= cachep
->allocflags
;
1877 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1878 flags
|= __GFP_RECLAIMABLE
;
1880 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1882 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1883 slab_out_of_memory(cachep
, flags
, nodeid
);
1887 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1888 if (unlikely(page
->pfmemalloc
))
1889 pfmemalloc_active
= true;
1891 nr_pages
= (1 << cachep
->gfporder
);
1892 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1893 add_zone_page_state(page_zone(page
),
1894 NR_SLAB_RECLAIMABLE
, nr_pages
);
1896 add_zone_page_state(page_zone(page
),
1897 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1898 for (i
= 0; i
< nr_pages
; i
++) {
1899 __SetPageSlab(page
+ i
);
1901 if (page
->pfmemalloc
)
1902 SetPageSlabPfmemalloc(page
+ i
);
1905 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1906 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1909 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1911 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1914 return page_address(page
);
1918 * Interface to system's page release.
1920 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1922 unsigned long i
= (1 << cachep
->gfporder
);
1923 struct page
*page
= virt_to_page(addr
);
1924 const unsigned long nr_freed
= i
;
1926 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1928 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1929 sub_zone_page_state(page_zone(page
),
1930 NR_SLAB_RECLAIMABLE
, nr_freed
);
1932 sub_zone_page_state(page_zone(page
),
1933 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1935 BUG_ON(!PageSlab(page
));
1936 __ClearPageSlabPfmemalloc(page
);
1937 __ClearPageSlab(page
);
1940 if (current
->reclaim_state
)
1941 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1942 free_pages((unsigned long)addr
, cachep
->gfporder
);
1945 static void kmem_rcu_free(struct rcu_head
*head
)
1947 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1948 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1950 kmem_freepages(cachep
, slab_rcu
->addr
);
1951 if (OFF_SLAB(cachep
))
1952 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1957 #ifdef CONFIG_DEBUG_PAGEALLOC
1958 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1959 unsigned long caller
)
1961 int size
= cachep
->object_size
;
1963 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1965 if (size
< 5 * sizeof(unsigned long))
1968 *addr
++ = 0x12345678;
1970 *addr
++ = smp_processor_id();
1971 size
-= 3 * sizeof(unsigned long);
1973 unsigned long *sptr
= &caller
;
1974 unsigned long svalue
;
1976 while (!kstack_end(sptr
)) {
1978 if (kernel_text_address(svalue
)) {
1980 size
-= sizeof(unsigned long);
1981 if (size
<= sizeof(unsigned long))
1987 *addr
++ = 0x87654321;
1991 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1993 int size
= cachep
->object_size
;
1994 addr
= &((char *)addr
)[obj_offset(cachep
)];
1996 memset(addr
, val
, size
);
1997 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
2000 static void dump_line(char *data
, int offset
, int limit
)
2003 unsigned char error
= 0;
2006 printk(KERN_ERR
"%03x: ", offset
);
2007 for (i
= 0; i
< limit
; i
++) {
2008 if (data
[offset
+ i
] != POISON_FREE
) {
2009 error
= data
[offset
+ i
];
2013 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
2014 &data
[offset
], limit
, 1);
2016 if (bad_count
== 1) {
2017 error
^= POISON_FREE
;
2018 if (!(error
& (error
- 1))) {
2019 printk(KERN_ERR
"Single bit error detected. Probably "
2022 printk(KERN_ERR
"Run memtest86+ or a similar memory "
2025 printk(KERN_ERR
"Run a memory test tool.\n");
2034 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
2039 if (cachep
->flags
& SLAB_RED_ZONE
) {
2040 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
2041 *dbg_redzone1(cachep
, objp
),
2042 *dbg_redzone2(cachep
, objp
));
2045 if (cachep
->flags
& SLAB_STORE_USER
) {
2046 printk(KERN_ERR
"Last user: [<%p>]",
2047 *dbg_userword(cachep
, objp
));
2048 print_symbol("(%s)",
2049 (unsigned long)*dbg_userword(cachep
, objp
));
2052 realobj
= (char *)objp
+ obj_offset(cachep
);
2053 size
= cachep
->object_size
;
2054 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
2057 if (i
+ limit
> size
)
2059 dump_line(realobj
, i
, limit
);
2063 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
2069 realobj
= (char *)objp
+ obj_offset(cachep
);
2070 size
= cachep
->object_size
;
2072 for (i
= 0; i
< size
; i
++) {
2073 char exp
= POISON_FREE
;
2076 if (realobj
[i
] != exp
) {
2082 "Slab corruption (%s): %s start=%p, len=%d\n",
2083 print_tainted(), cachep
->name
, realobj
, size
);
2084 print_objinfo(cachep
, objp
, 0);
2086 /* Hexdump the affected line */
2089 if (i
+ limit
> size
)
2091 dump_line(realobj
, i
, limit
);
2094 /* Limit to 5 lines */
2100 /* Print some data about the neighboring objects, if they
2103 struct slab
*slabp
= virt_to_slab(objp
);
2106 objnr
= obj_to_index(cachep
, slabp
, objp
);
2108 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
2109 realobj
= (char *)objp
+ obj_offset(cachep
);
2110 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
2112 print_objinfo(cachep
, objp
, 2);
2114 if (objnr
+ 1 < cachep
->num
) {
2115 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
2116 realobj
= (char *)objp
+ obj_offset(cachep
);
2117 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
2119 print_objinfo(cachep
, objp
, 2);
2126 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2129 for (i
= 0; i
< cachep
->num
; i
++) {
2130 void *objp
= index_to_obj(cachep
, slabp
, i
);
2132 if (cachep
->flags
& SLAB_POISON
) {
2133 #ifdef CONFIG_DEBUG_PAGEALLOC
2134 if (cachep
->size
% PAGE_SIZE
== 0 &&
2136 kernel_map_pages(virt_to_page(objp
),
2137 cachep
->size
/ PAGE_SIZE
, 1);
2139 check_poison_obj(cachep
, objp
);
2141 check_poison_obj(cachep
, objp
);
2144 if (cachep
->flags
& SLAB_RED_ZONE
) {
2145 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2146 slab_error(cachep
, "start of a freed object "
2148 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2149 slab_error(cachep
, "end of a freed object "
2155 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2161 * slab_destroy - destroy and release all objects in a slab
2162 * @cachep: cache pointer being destroyed
2163 * @slabp: slab pointer being destroyed
2165 * Destroy all the objs in a slab, and release the mem back to the system.
2166 * Before calling the slab must have been unlinked from the cache. The
2167 * cache-lock is not held/needed.
2169 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2171 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2173 slab_destroy_debugcheck(cachep
, slabp
);
2174 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2175 struct slab_rcu
*slab_rcu
;
2177 slab_rcu
= (struct slab_rcu
*)slabp
;
2178 slab_rcu
->cachep
= cachep
;
2179 slab_rcu
->addr
= addr
;
2180 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2182 kmem_freepages(cachep
, addr
);
2183 if (OFF_SLAB(cachep
))
2184 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2188 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
2191 struct kmem_list3
*l3
;
2193 for_each_online_cpu(i
)
2194 kfree(cachep
->array
[i
]);
2196 /* NUMA: free the list3 structures */
2197 for_each_online_node(i
) {
2198 l3
= cachep
->nodelists
[i
];
2201 free_alien_cache(l3
->alien
);
2205 kmem_cache_free(&cache_cache
, cachep
);
2210 * calculate_slab_order - calculate size (page order) of slabs
2211 * @cachep: pointer to the cache that is being created
2212 * @size: size of objects to be created in this cache.
2213 * @align: required alignment for the objects.
2214 * @flags: slab allocation flags
2216 * Also calculates the number of objects per slab.
2218 * This could be made much more intelligent. For now, try to avoid using
2219 * high order pages for slabs. When the gfp() functions are more friendly
2220 * towards high-order requests, this should be changed.
2222 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2223 size_t size
, size_t align
, unsigned long flags
)
2225 unsigned long offslab_limit
;
2226 size_t left_over
= 0;
2229 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2233 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2237 if (flags
& CFLGS_OFF_SLAB
) {
2239 * Max number of objs-per-slab for caches which
2240 * use off-slab slabs. Needed to avoid a possible
2241 * looping condition in cache_grow().
2243 offslab_limit
= size
- sizeof(struct slab
);
2244 offslab_limit
/= sizeof(kmem_bufctl_t
);
2246 if (num
> offslab_limit
)
2250 /* Found something acceptable - save it away */
2252 cachep
->gfporder
= gfporder
;
2253 left_over
= remainder
;
2256 * A VFS-reclaimable slab tends to have most allocations
2257 * as GFP_NOFS and we really don't want to have to be allocating
2258 * higher-order pages when we are unable to shrink dcache.
2260 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2264 * Large number of objects is good, but very large slabs are
2265 * currently bad for the gfp()s.
2267 if (gfporder
>= slab_max_order
)
2271 * Acceptable internal fragmentation?
2273 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2279 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2281 if (slab_state
>= FULL
)
2282 return enable_cpucache(cachep
, gfp
);
2284 if (slab_state
== DOWN
) {
2286 * Note: the first kmem_cache_create must create the cache
2287 * that's used by kmalloc(24), otherwise the creation of
2288 * further caches will BUG().
2290 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2293 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2294 * the first cache, then we need to set up all its list3s,
2295 * otherwise the creation of further caches will BUG().
2297 set_up_list3s(cachep
, SIZE_AC
);
2298 if (INDEX_AC
== INDEX_L3
)
2299 slab_state
= PARTIAL_L3
;
2301 slab_state
= PARTIAL_ARRAYCACHE
;
2303 cachep
->array
[smp_processor_id()] =
2304 kmalloc(sizeof(struct arraycache_init
), gfp
);
2306 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2307 set_up_list3s(cachep
, SIZE_L3
);
2308 slab_state
= PARTIAL_L3
;
2311 for_each_online_node(node
) {
2312 cachep
->nodelists
[node
] =
2313 kmalloc_node(sizeof(struct kmem_list3
),
2315 BUG_ON(!cachep
->nodelists
[node
]);
2316 kmem_list3_init(cachep
->nodelists
[node
]);
2320 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2321 jiffies
+ REAPTIMEOUT_LIST3
+
2322 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2324 cpu_cache_get(cachep
)->avail
= 0;
2325 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2326 cpu_cache_get(cachep
)->batchcount
= 1;
2327 cpu_cache_get(cachep
)->touched
= 0;
2328 cachep
->batchcount
= 1;
2329 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2334 * __kmem_cache_create - Create a cache.
2335 * @name: A string which is used in /proc/slabinfo to identify this cache.
2336 * @size: The size of objects to be created in this cache.
2337 * @align: The required alignment for the objects.
2338 * @flags: SLAB flags
2339 * @ctor: A constructor for the objects.
2341 * Returns a ptr to the cache on success, NULL on failure.
2342 * Cannot be called within a int, but can be interrupted.
2343 * The @ctor is run when new pages are allocated by the cache.
2345 * @name must be valid until the cache is destroyed. This implies that
2346 * the module calling this has to destroy the cache before getting unloaded.
2350 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2351 * to catch references to uninitialised memory.
2353 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2354 * for buffer overruns.
2356 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2357 * cacheline. This can be beneficial if you're counting cycles as closely
2361 __kmem_cache_create (const char *name
, size_t size
, size_t align
,
2362 unsigned long flags
, void (*ctor
)(void *))
2364 size_t left_over
, slab_size
, ralign
;
2365 struct kmem_cache
*cachep
= NULL
;
2371 * Enable redzoning and last user accounting, except for caches with
2372 * large objects, if the increased size would increase the object size
2373 * above the next power of two: caches with object sizes just above a
2374 * power of two have a significant amount of internal fragmentation.
2376 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2377 2 * sizeof(unsigned long long)))
2378 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2379 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2380 flags
|= SLAB_POISON
;
2382 if (flags
& SLAB_DESTROY_BY_RCU
)
2383 BUG_ON(flags
& SLAB_POISON
);
2386 * Always checks flags, a caller might be expecting debug support which
2389 BUG_ON(flags
& ~CREATE_MASK
);
2392 * Check that size is in terms of words. This is needed to avoid
2393 * unaligned accesses for some archs when redzoning is used, and makes
2394 * sure any on-slab bufctl's are also correctly aligned.
2396 if (size
& (BYTES_PER_WORD
- 1)) {
2397 size
+= (BYTES_PER_WORD
- 1);
2398 size
&= ~(BYTES_PER_WORD
- 1);
2401 /* calculate the final buffer alignment: */
2403 /* 1) arch recommendation: can be overridden for debug */
2404 if (flags
& SLAB_HWCACHE_ALIGN
) {
2406 * Default alignment: as specified by the arch code. Except if
2407 * an object is really small, then squeeze multiple objects into
2410 ralign
= cache_line_size();
2411 while (size
<= ralign
/ 2)
2414 ralign
= BYTES_PER_WORD
;
2418 * Redzoning and user store require word alignment or possibly larger.
2419 * Note this will be overridden by architecture or caller mandated
2420 * alignment if either is greater than BYTES_PER_WORD.
2422 if (flags
& SLAB_STORE_USER
)
2423 ralign
= BYTES_PER_WORD
;
2425 if (flags
& SLAB_RED_ZONE
) {
2426 ralign
= REDZONE_ALIGN
;
2427 /* If redzoning, ensure that the second redzone is suitably
2428 * aligned, by adjusting the object size accordingly. */
2429 size
+= REDZONE_ALIGN
- 1;
2430 size
&= ~(REDZONE_ALIGN
- 1);
2433 /* 2) arch mandated alignment */
2434 if (ralign
< ARCH_SLAB_MINALIGN
) {
2435 ralign
= ARCH_SLAB_MINALIGN
;
2437 /* 3) caller mandated alignment */
2438 if (ralign
< align
) {
2441 /* disable debug if necessary */
2442 if (ralign
> __alignof__(unsigned long long))
2443 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2449 if (slab_is_available())
2454 /* Get cache's description obj. */
2455 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2459 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
2460 cachep
->object_size
= size
;
2461 cachep
->align
= align
;
2465 * Both debugging options require word-alignment which is calculated
2468 if (flags
& SLAB_RED_ZONE
) {
2469 /* add space for red zone words */
2470 cachep
->obj_offset
+= sizeof(unsigned long long);
2471 size
+= 2 * sizeof(unsigned long long);
2473 if (flags
& SLAB_STORE_USER
) {
2474 /* user store requires one word storage behind the end of
2475 * the real object. But if the second red zone needs to be
2476 * aligned to 64 bits, we must allow that much space.
2478 if (flags
& SLAB_RED_ZONE
)
2479 size
+= REDZONE_ALIGN
;
2481 size
+= BYTES_PER_WORD
;
2483 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2484 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2485 && cachep
->object_size
> cache_line_size() && ALIGN(size
, align
) < PAGE_SIZE
) {
2486 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, align
);
2493 * Determine if the slab management is 'on' or 'off' slab.
2494 * (bootstrapping cannot cope with offslab caches so don't do
2495 * it too early on. Always use on-slab management when
2496 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2498 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2499 !(flags
& SLAB_NOLEAKTRACE
))
2501 * Size is large, assume best to place the slab management obj
2502 * off-slab (should allow better packing of objs).
2504 flags
|= CFLGS_OFF_SLAB
;
2506 size
= ALIGN(size
, align
);
2508 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2512 "kmem_cache_create: couldn't create cache %s.\n", name
);
2513 kmem_cache_free(&cache_cache
, cachep
);
2516 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2517 + sizeof(struct slab
), align
);
2520 * If the slab has been placed off-slab, and we have enough space then
2521 * move it on-slab. This is at the expense of any extra colouring.
2523 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2524 flags
&= ~CFLGS_OFF_SLAB
;
2525 left_over
-= slab_size
;
2528 if (flags
& CFLGS_OFF_SLAB
) {
2529 /* really off slab. No need for manual alignment */
2531 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2533 #ifdef CONFIG_PAGE_POISONING
2534 /* If we're going to use the generic kernel_map_pages()
2535 * poisoning, then it's going to smash the contents of
2536 * the redzone and userword anyhow, so switch them off.
2538 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2539 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2543 cachep
->colour_off
= cache_line_size();
2544 /* Offset must be a multiple of the alignment. */
2545 if (cachep
->colour_off
< align
)
2546 cachep
->colour_off
= align
;
2547 cachep
->colour
= left_over
/ cachep
->colour_off
;
2548 cachep
->slab_size
= slab_size
;
2549 cachep
->flags
= flags
;
2550 cachep
->allocflags
= 0;
2551 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2552 cachep
->allocflags
|= GFP_DMA
;
2553 cachep
->size
= size
;
2554 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2556 if (flags
& CFLGS_OFF_SLAB
) {
2557 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2559 * This is a possibility for one of the malloc_sizes caches.
2560 * But since we go off slab only for object size greater than
2561 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2562 * this should not happen at all.
2563 * But leave a BUG_ON for some lucky dude.
2565 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2567 cachep
->ctor
= ctor
;
2568 cachep
->name
= name
;
2570 if (setup_cpu_cache(cachep
, gfp
)) {
2571 __kmem_cache_destroy(cachep
);
2575 if (flags
& SLAB_DEBUG_OBJECTS
) {
2577 * Would deadlock through slab_destroy()->call_rcu()->
2578 * debug_object_activate()->kmem_cache_alloc().
2580 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2582 slab_set_debugobj_lock_classes(cachep
);
2585 /* cache setup completed, link it into the list */
2586 list_add(&cachep
->list
, &slab_caches
);
2591 static void check_irq_off(void)
2593 BUG_ON(!irqs_disabled());
2596 static void check_irq_on(void)
2598 BUG_ON(irqs_disabled());
2601 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2605 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2609 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2613 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2618 #define check_irq_off() do { } while(0)
2619 #define check_irq_on() do { } while(0)
2620 #define check_spinlock_acquired(x) do { } while(0)
2621 #define check_spinlock_acquired_node(x, y) do { } while(0)
2624 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2625 struct array_cache
*ac
,
2626 int force
, int node
);
2628 static void do_drain(void *arg
)
2630 struct kmem_cache
*cachep
= arg
;
2631 struct array_cache
*ac
;
2632 int node
= numa_mem_id();
2635 ac
= cpu_cache_get(cachep
);
2636 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2637 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2638 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2642 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2644 struct kmem_list3
*l3
;
2647 on_each_cpu(do_drain
, cachep
, 1);
2649 for_each_online_node(node
) {
2650 l3
= cachep
->nodelists
[node
];
2651 if (l3
&& l3
->alien
)
2652 drain_alien_cache(cachep
, l3
->alien
);
2655 for_each_online_node(node
) {
2656 l3
= cachep
->nodelists
[node
];
2658 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2663 * Remove slabs from the list of free slabs.
2664 * Specify the number of slabs to drain in tofree.
2666 * Returns the actual number of slabs released.
2668 static int drain_freelist(struct kmem_cache
*cache
,
2669 struct kmem_list3
*l3
, int tofree
)
2671 struct list_head
*p
;
2676 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2678 spin_lock_irq(&l3
->list_lock
);
2679 p
= l3
->slabs_free
.prev
;
2680 if (p
== &l3
->slabs_free
) {
2681 spin_unlock_irq(&l3
->list_lock
);
2685 slabp
= list_entry(p
, struct slab
, list
);
2687 BUG_ON(slabp
->inuse
);
2689 list_del(&slabp
->list
);
2691 * Safe to drop the lock. The slab is no longer linked
2694 l3
->free_objects
-= cache
->num
;
2695 spin_unlock_irq(&l3
->list_lock
);
2696 slab_destroy(cache
, slabp
);
2703 /* Called with slab_mutex held to protect against cpu hotplug */
2704 static int __cache_shrink(struct kmem_cache
*cachep
)
2707 struct kmem_list3
*l3
;
2709 drain_cpu_caches(cachep
);
2712 for_each_online_node(i
) {
2713 l3
= cachep
->nodelists
[i
];
2717 drain_freelist(cachep
, l3
, l3
->free_objects
);
2719 ret
+= !list_empty(&l3
->slabs_full
) ||
2720 !list_empty(&l3
->slabs_partial
);
2722 return (ret
? 1 : 0);
2726 * kmem_cache_shrink - Shrink a cache.
2727 * @cachep: The cache to shrink.
2729 * Releases as many slabs as possible for a cache.
2730 * To help debugging, a zero exit status indicates all slabs were released.
2732 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2735 BUG_ON(!cachep
|| in_interrupt());
2738 mutex_lock(&slab_mutex
);
2739 ret
= __cache_shrink(cachep
);
2740 mutex_unlock(&slab_mutex
);
2744 EXPORT_SYMBOL(kmem_cache_shrink
);
2747 * kmem_cache_destroy - delete a cache
2748 * @cachep: the cache to destroy
2750 * Remove a &struct kmem_cache object from the slab cache.
2752 * It is expected this function will be called by a module when it is
2753 * unloaded. This will remove the cache completely, and avoid a duplicate
2754 * cache being allocated each time a module is loaded and unloaded, if the
2755 * module doesn't have persistent in-kernel storage across loads and unloads.
2757 * The cache must be empty before calling this function.
2759 * The caller must guarantee that no one will allocate memory from the cache
2760 * during the kmem_cache_destroy().
2762 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2764 BUG_ON(!cachep
|| in_interrupt());
2766 /* Find the cache in the chain of caches. */
2768 mutex_lock(&slab_mutex
);
2770 * the chain is never empty, cache_cache is never destroyed
2772 list_del(&cachep
->list
);
2773 if (__cache_shrink(cachep
)) {
2774 slab_error(cachep
, "Can't free all objects");
2775 list_add(&cachep
->list
, &slab_caches
);
2776 mutex_unlock(&slab_mutex
);
2781 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2784 __kmem_cache_destroy(cachep
);
2785 mutex_unlock(&slab_mutex
);
2788 EXPORT_SYMBOL(kmem_cache_destroy
);
2791 * Get the memory for a slab management obj.
2792 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2793 * always come from malloc_sizes caches. The slab descriptor cannot
2794 * come from the same cache which is getting created because,
2795 * when we are searching for an appropriate cache for these
2796 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2797 * If we are creating a malloc_sizes cache here it would not be visible to
2798 * kmem_find_general_cachep till the initialization is complete.
2799 * Hence we cannot have slabp_cache same as the original cache.
2801 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2802 int colour_off
, gfp_t local_flags
,
2807 if (OFF_SLAB(cachep
)) {
2808 /* Slab management obj is off-slab. */
2809 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2810 local_flags
, nodeid
);
2812 * If the first object in the slab is leaked (it's allocated
2813 * but no one has a reference to it), we want to make sure
2814 * kmemleak does not treat the ->s_mem pointer as a reference
2815 * to the object. Otherwise we will not report the leak.
2817 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2822 slabp
= objp
+ colour_off
;
2823 colour_off
+= cachep
->slab_size
;
2826 slabp
->colouroff
= colour_off
;
2827 slabp
->s_mem
= objp
+ colour_off
;
2828 slabp
->nodeid
= nodeid
;
2833 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2835 return (kmem_bufctl_t
*) (slabp
+ 1);
2838 static void cache_init_objs(struct kmem_cache
*cachep
,
2843 for (i
= 0; i
< cachep
->num
; i
++) {
2844 void *objp
= index_to_obj(cachep
, slabp
, i
);
2846 /* need to poison the objs? */
2847 if (cachep
->flags
& SLAB_POISON
)
2848 poison_obj(cachep
, objp
, POISON_FREE
);
2849 if (cachep
->flags
& SLAB_STORE_USER
)
2850 *dbg_userword(cachep
, objp
) = NULL
;
2852 if (cachep
->flags
& SLAB_RED_ZONE
) {
2853 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2854 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2857 * Constructors are not allowed to allocate memory from the same
2858 * cache which they are a constructor for. Otherwise, deadlock.
2859 * They must also be threaded.
2861 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2862 cachep
->ctor(objp
+ obj_offset(cachep
));
2864 if (cachep
->flags
& SLAB_RED_ZONE
) {
2865 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2866 slab_error(cachep
, "constructor overwrote the"
2867 " end of an object");
2868 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2869 slab_error(cachep
, "constructor overwrote the"
2870 " start of an object");
2872 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2873 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2874 kernel_map_pages(virt_to_page(objp
),
2875 cachep
->size
/ PAGE_SIZE
, 0);
2880 slab_bufctl(slabp
)[i
] = i
+ 1;
2882 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2885 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2887 if (CONFIG_ZONE_DMA_FLAG
) {
2888 if (flags
& GFP_DMA
)
2889 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2891 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2895 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2898 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2902 next
= slab_bufctl(slabp
)[slabp
->free
];
2904 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2905 WARN_ON(slabp
->nodeid
!= nodeid
);
2912 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2913 void *objp
, int nodeid
)
2915 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2918 /* Verify that the slab belongs to the intended node */
2919 WARN_ON(slabp
->nodeid
!= nodeid
);
2921 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2922 printk(KERN_ERR
"slab: double free detected in cache "
2923 "'%s', objp %p\n", cachep
->name
, objp
);
2927 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2928 slabp
->free
= objnr
;
2933 * Map pages beginning at addr to the given cache and slab. This is required
2934 * for the slab allocator to be able to lookup the cache and slab of a
2935 * virtual address for kfree, ksize, and slab debugging.
2937 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2943 page
= virt_to_page(addr
);
2946 if (likely(!PageCompound(page
)))
2947 nr_pages
<<= cache
->gfporder
;
2950 page
->slab_cache
= cache
;
2951 page
->slab_page
= slab
;
2953 } while (--nr_pages
);
2957 * Grow (by 1) the number of slabs within a cache. This is called by
2958 * kmem_cache_alloc() when there are no active objs left in a cache.
2960 static int cache_grow(struct kmem_cache
*cachep
,
2961 gfp_t flags
, int nodeid
, void *objp
)
2966 struct kmem_list3
*l3
;
2969 * Be lazy and only check for valid flags here, keeping it out of the
2970 * critical path in kmem_cache_alloc().
2972 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2973 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2975 /* Take the l3 list lock to change the colour_next on this node */
2977 l3
= cachep
->nodelists
[nodeid
];
2978 spin_lock(&l3
->list_lock
);
2980 /* Get colour for the slab, and cal the next value. */
2981 offset
= l3
->colour_next
;
2983 if (l3
->colour_next
>= cachep
->colour
)
2984 l3
->colour_next
= 0;
2985 spin_unlock(&l3
->list_lock
);
2987 offset
*= cachep
->colour_off
;
2989 if (local_flags
& __GFP_WAIT
)
2993 * The test for missing atomic flag is performed here, rather than
2994 * the more obvious place, simply to reduce the critical path length
2995 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2996 * will eventually be caught here (where it matters).
2998 kmem_flagcheck(cachep
, flags
);
3001 * Get mem for the objs. Attempt to allocate a physical page from
3005 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
3009 /* Get slab management. */
3010 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
3011 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
3015 slab_map_pages(cachep
, slabp
, objp
);
3017 cache_init_objs(cachep
, slabp
);
3019 if (local_flags
& __GFP_WAIT
)
3020 local_irq_disable();
3022 spin_lock(&l3
->list_lock
);
3024 /* Make slab active. */
3025 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
3026 STATS_INC_GROWN(cachep
);
3027 l3
->free_objects
+= cachep
->num
;
3028 spin_unlock(&l3
->list_lock
);
3031 kmem_freepages(cachep
, objp
);
3033 if (local_flags
& __GFP_WAIT
)
3034 local_irq_disable();
3041 * Perform extra freeing checks:
3042 * - detect bad pointers.
3043 * - POISON/RED_ZONE checking
3045 static void kfree_debugcheck(const void *objp
)
3047 if (!virt_addr_valid(objp
)) {
3048 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
3049 (unsigned long)objp
);
3054 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
3056 unsigned long long redzone1
, redzone2
;
3058 redzone1
= *dbg_redzone1(cache
, obj
);
3059 redzone2
= *dbg_redzone2(cache
, obj
);
3064 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
3067 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
3068 slab_error(cache
, "double free detected");
3070 slab_error(cache
, "memory outside object was overwritten");
3072 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3073 obj
, redzone1
, redzone2
);
3076 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
3083 BUG_ON(virt_to_cache(objp
) != cachep
);
3085 objp
-= obj_offset(cachep
);
3086 kfree_debugcheck(objp
);
3087 page
= virt_to_head_page(objp
);
3089 slabp
= page
->slab_page
;
3091 if (cachep
->flags
& SLAB_RED_ZONE
) {
3092 verify_redzone_free(cachep
, objp
);
3093 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
3094 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
3096 if (cachep
->flags
& SLAB_STORE_USER
)
3097 *dbg_userword(cachep
, objp
) = caller
;
3099 objnr
= obj_to_index(cachep
, slabp
, objp
);
3101 BUG_ON(objnr
>= cachep
->num
);
3102 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
3104 #ifdef CONFIG_DEBUG_SLAB_LEAK
3105 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
3107 if (cachep
->flags
& SLAB_POISON
) {
3108 #ifdef CONFIG_DEBUG_PAGEALLOC
3109 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
3110 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
3111 kernel_map_pages(virt_to_page(objp
),
3112 cachep
->size
/ PAGE_SIZE
, 0);
3114 poison_obj(cachep
, objp
, POISON_FREE
);
3117 poison_obj(cachep
, objp
, POISON_FREE
);
3123 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
3128 /* Check slab's freelist to see if this obj is there. */
3129 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
3131 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3134 if (entries
!= cachep
->num
- slabp
->inuse
) {
3136 printk(KERN_ERR
"slab: Internal list corruption detected in "
3137 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3138 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
3140 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3141 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3147 #define kfree_debugcheck(x) do { } while(0)
3148 #define cache_free_debugcheck(x,objp,z) (objp)
3149 #define check_slabp(x,y) do { } while(0)
3152 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
3156 struct kmem_list3
*l3
;
3157 struct array_cache
*ac
;
3161 node
= numa_mem_id();
3162 if (unlikely(force_refill
))
3165 ac
= cpu_cache_get(cachep
);
3166 batchcount
= ac
->batchcount
;
3167 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3169 * If there was little recent activity on this cache, then
3170 * perform only a partial refill. Otherwise we could generate
3173 batchcount
= BATCHREFILL_LIMIT
;
3175 l3
= cachep
->nodelists
[node
];
3177 BUG_ON(ac
->avail
> 0 || !l3
);
3178 spin_lock(&l3
->list_lock
);
3180 /* See if we can refill from the shared array */
3181 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3182 l3
->shared
->touched
= 1;
3186 while (batchcount
> 0) {
3187 struct list_head
*entry
;
3189 /* Get slab alloc is to come from. */
3190 entry
= l3
->slabs_partial
.next
;
3191 if (entry
== &l3
->slabs_partial
) {
3192 l3
->free_touched
= 1;
3193 entry
= l3
->slabs_free
.next
;
3194 if (entry
== &l3
->slabs_free
)
3198 slabp
= list_entry(entry
, struct slab
, list
);
3199 check_slabp(cachep
, slabp
);
3200 check_spinlock_acquired(cachep
);
3203 * The slab was either on partial or free list so
3204 * there must be at least one object available for
3207 BUG_ON(slabp
->inuse
>= cachep
->num
);
3209 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3210 STATS_INC_ALLOCED(cachep
);
3211 STATS_INC_ACTIVE(cachep
);
3212 STATS_SET_HIGH(cachep
);
3214 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
3217 check_slabp(cachep
, slabp
);
3219 /* move slabp to correct slabp list: */
3220 list_del(&slabp
->list
);
3221 if (slabp
->free
== BUFCTL_END
)
3222 list_add(&slabp
->list
, &l3
->slabs_full
);
3224 list_add(&slabp
->list
, &l3
->slabs_partial
);
3228 l3
->free_objects
-= ac
->avail
;
3230 spin_unlock(&l3
->list_lock
);
3232 if (unlikely(!ac
->avail
)) {
3235 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3237 /* cache_grow can reenable interrupts, then ac could change. */
3238 ac
= cpu_cache_get(cachep
);
3240 /* no objects in sight? abort */
3241 if (!x
&& (ac
->avail
== 0 || force_refill
))
3244 if (!ac
->avail
) /* objects refilled by interrupt? */
3249 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
3252 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3255 might_sleep_if(flags
& __GFP_WAIT
);
3257 kmem_flagcheck(cachep
, flags
);
3262 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3263 gfp_t flags
, void *objp
, void *caller
)
3267 if (cachep
->flags
& SLAB_POISON
) {
3268 #ifdef CONFIG_DEBUG_PAGEALLOC
3269 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3270 kernel_map_pages(virt_to_page(objp
),
3271 cachep
->size
/ PAGE_SIZE
, 1);
3273 check_poison_obj(cachep
, objp
);
3275 check_poison_obj(cachep
, objp
);
3277 poison_obj(cachep
, objp
, POISON_INUSE
);
3279 if (cachep
->flags
& SLAB_STORE_USER
)
3280 *dbg_userword(cachep
, objp
) = caller
;
3282 if (cachep
->flags
& SLAB_RED_ZONE
) {
3283 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3284 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3285 slab_error(cachep
, "double free, or memory outside"
3286 " object was overwritten");
3288 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3289 objp
, *dbg_redzone1(cachep
, objp
),
3290 *dbg_redzone2(cachep
, objp
));
3292 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3293 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3295 #ifdef CONFIG_DEBUG_SLAB_LEAK
3300 slabp
= virt_to_head_page(objp
)->slab_page
;
3301 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3302 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3305 objp
+= obj_offset(cachep
);
3306 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3308 if (ARCH_SLAB_MINALIGN
&&
3309 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3310 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3311 objp
, (int)ARCH_SLAB_MINALIGN
);
3316 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3319 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3321 if (cachep
== &cache_cache
)
3324 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3327 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3330 struct array_cache
*ac
;
3331 bool force_refill
= false;
3335 ac
= cpu_cache_get(cachep
);
3336 if (likely(ac
->avail
)) {
3338 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3341 * Allow for the possibility all avail objects are not allowed
3342 * by the current flags
3345 STATS_INC_ALLOCHIT(cachep
);
3348 force_refill
= true;
3351 STATS_INC_ALLOCMISS(cachep
);
3352 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3354 * the 'ac' may be updated by cache_alloc_refill(),
3355 * and kmemleak_erase() requires its correct value.
3357 ac
= cpu_cache_get(cachep
);
3361 * To avoid a false negative, if an object that is in one of the
3362 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3363 * treat the array pointers as a reference to the object.
3366 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3372 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3374 * If we are in_interrupt, then process context, including cpusets and
3375 * mempolicy, may not apply and should not be used for allocation policy.
3377 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3379 int nid_alloc
, nid_here
;
3381 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3383 nid_alloc
= nid_here
= numa_mem_id();
3384 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3385 nid_alloc
= cpuset_slab_spread_node();
3386 else if (current
->mempolicy
)
3387 nid_alloc
= slab_node();
3388 if (nid_alloc
!= nid_here
)
3389 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3394 * Fallback function if there was no memory available and no objects on a
3395 * certain node and fall back is permitted. First we scan all the
3396 * available nodelists for available objects. If that fails then we
3397 * perform an allocation without specifying a node. This allows the page
3398 * allocator to do its reclaim / fallback magic. We then insert the
3399 * slab into the proper nodelist and then allocate from it.
3401 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3403 struct zonelist
*zonelist
;
3407 enum zone_type high_zoneidx
= gfp_zone(flags
);
3410 unsigned int cpuset_mems_cookie
;
3412 if (flags
& __GFP_THISNODE
)
3415 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3418 cpuset_mems_cookie
= get_mems_allowed();
3419 zonelist
= node_zonelist(slab_node(), flags
);
3423 * Look through allowed nodes for objects available
3424 * from existing per node queues.
3426 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3427 nid
= zone_to_nid(zone
);
3429 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3430 cache
->nodelists
[nid
] &&
3431 cache
->nodelists
[nid
]->free_objects
) {
3432 obj
= ____cache_alloc_node(cache
,
3433 flags
| GFP_THISNODE
, nid
);
3441 * This allocation will be performed within the constraints
3442 * of the current cpuset / memory policy requirements.
3443 * We may trigger various forms of reclaim on the allowed
3444 * set and go into memory reserves if necessary.
3446 if (local_flags
& __GFP_WAIT
)
3448 kmem_flagcheck(cache
, flags
);
3449 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3450 if (local_flags
& __GFP_WAIT
)
3451 local_irq_disable();
3454 * Insert into the appropriate per node queues
3456 nid
= page_to_nid(virt_to_page(obj
));
3457 if (cache_grow(cache
, flags
, nid
, obj
)) {
3458 obj
= ____cache_alloc_node(cache
,
3459 flags
| GFP_THISNODE
, nid
);
3462 * Another processor may allocate the
3463 * objects in the slab since we are
3464 * not holding any locks.
3468 /* cache_grow already freed obj */
3474 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3480 * A interface to enable slab creation on nodeid
3482 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3485 struct list_head
*entry
;
3487 struct kmem_list3
*l3
;
3491 l3
= cachep
->nodelists
[nodeid
];
3496 spin_lock(&l3
->list_lock
);
3497 entry
= l3
->slabs_partial
.next
;
3498 if (entry
== &l3
->slabs_partial
) {
3499 l3
->free_touched
= 1;
3500 entry
= l3
->slabs_free
.next
;
3501 if (entry
== &l3
->slabs_free
)
3505 slabp
= list_entry(entry
, struct slab
, list
);
3506 check_spinlock_acquired_node(cachep
, nodeid
);
3507 check_slabp(cachep
, slabp
);
3509 STATS_INC_NODEALLOCS(cachep
);
3510 STATS_INC_ACTIVE(cachep
);
3511 STATS_SET_HIGH(cachep
);
3513 BUG_ON(slabp
->inuse
== cachep
->num
);
3515 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3516 check_slabp(cachep
, slabp
);
3518 /* move slabp to correct slabp list: */
3519 list_del(&slabp
->list
);
3521 if (slabp
->free
== BUFCTL_END
)
3522 list_add(&slabp
->list
, &l3
->slabs_full
);
3524 list_add(&slabp
->list
, &l3
->slabs_partial
);
3526 spin_unlock(&l3
->list_lock
);
3530 spin_unlock(&l3
->list_lock
);
3531 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3535 return fallback_alloc(cachep
, flags
);
3542 * kmem_cache_alloc_node - Allocate an object on the specified node
3543 * @cachep: The cache to allocate from.
3544 * @flags: See kmalloc().
3545 * @nodeid: node number of the target node.
3546 * @caller: return address of caller, used for debug information
3548 * Identical to kmem_cache_alloc but it will allocate memory on the given
3549 * node, which can improve the performance for cpu bound structures.
3551 * Fallback to other node is possible if __GFP_THISNODE is not set.
3553 static __always_inline
void *
3554 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3557 unsigned long save_flags
;
3559 int slab_node
= numa_mem_id();
3561 flags
&= gfp_allowed_mask
;
3563 lockdep_trace_alloc(flags
);
3565 if (slab_should_failslab(cachep
, flags
))
3568 cache_alloc_debugcheck_before(cachep
, flags
);
3569 local_irq_save(save_flags
);
3571 if (nodeid
== NUMA_NO_NODE
)
3574 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3575 /* Node not bootstrapped yet */
3576 ptr
= fallback_alloc(cachep
, flags
);
3580 if (nodeid
== slab_node
) {
3582 * Use the locally cached objects if possible.
3583 * However ____cache_alloc does not allow fallback
3584 * to other nodes. It may fail while we still have
3585 * objects on other nodes available.
3587 ptr
= ____cache_alloc(cachep
, flags
);
3591 /* ___cache_alloc_node can fall back to other nodes */
3592 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3594 local_irq_restore(save_flags
);
3595 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3596 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3600 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3602 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3603 memset(ptr
, 0, cachep
->object_size
);
3608 static __always_inline
void *
3609 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3613 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3614 objp
= alternate_node_alloc(cache
, flags
);
3618 objp
= ____cache_alloc(cache
, flags
);
3621 * We may just have run out of memory on the local node.
3622 * ____cache_alloc_node() knows how to locate memory on other nodes
3625 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3632 static __always_inline
void *
3633 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3635 return ____cache_alloc(cachep
, flags
);
3638 #endif /* CONFIG_NUMA */
3640 static __always_inline
void *
3641 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3643 unsigned long save_flags
;
3646 flags
&= gfp_allowed_mask
;
3648 lockdep_trace_alloc(flags
);
3650 if (slab_should_failslab(cachep
, flags
))
3653 cache_alloc_debugcheck_before(cachep
, flags
);
3654 local_irq_save(save_flags
);
3655 objp
= __do_cache_alloc(cachep
, flags
);
3656 local_irq_restore(save_flags
);
3657 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3658 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3663 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3665 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3666 memset(objp
, 0, cachep
->object_size
);
3672 * Caller needs to acquire correct kmem_list's list_lock
3674 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3678 struct kmem_list3
*l3
;
3680 for (i
= 0; i
< nr_objects
; i
++) {
3684 clear_obj_pfmemalloc(&objpp
[i
]);
3687 slabp
= virt_to_slab(objp
);
3688 l3
= cachep
->nodelists
[node
];
3689 list_del(&slabp
->list
);
3690 check_spinlock_acquired_node(cachep
, node
);
3691 check_slabp(cachep
, slabp
);
3692 slab_put_obj(cachep
, slabp
, objp
, node
);
3693 STATS_DEC_ACTIVE(cachep
);
3695 check_slabp(cachep
, slabp
);
3697 /* fixup slab chains */
3698 if (slabp
->inuse
== 0) {
3699 if (l3
->free_objects
> l3
->free_limit
) {
3700 l3
->free_objects
-= cachep
->num
;
3701 /* No need to drop any previously held
3702 * lock here, even if we have a off-slab slab
3703 * descriptor it is guaranteed to come from
3704 * a different cache, refer to comments before
3707 slab_destroy(cachep
, slabp
);
3709 list_add(&slabp
->list
, &l3
->slabs_free
);
3712 /* Unconditionally move a slab to the end of the
3713 * partial list on free - maximum time for the
3714 * other objects to be freed, too.
3716 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3721 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3724 struct kmem_list3
*l3
;
3725 int node
= numa_mem_id();
3727 batchcount
= ac
->batchcount
;
3729 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3732 l3
= cachep
->nodelists
[node
];
3733 spin_lock(&l3
->list_lock
);
3735 struct array_cache
*shared_array
= l3
->shared
;
3736 int max
= shared_array
->limit
- shared_array
->avail
;
3738 if (batchcount
> max
)
3740 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3741 ac
->entry
, sizeof(void *) * batchcount
);
3742 shared_array
->avail
+= batchcount
;
3747 free_block(cachep
, ac
->entry
, batchcount
, node
);
3752 struct list_head
*p
;
3754 p
= l3
->slabs_free
.next
;
3755 while (p
!= &(l3
->slabs_free
)) {
3758 slabp
= list_entry(p
, struct slab
, list
);
3759 BUG_ON(slabp
->inuse
);
3764 STATS_SET_FREEABLE(cachep
, i
);
3767 spin_unlock(&l3
->list_lock
);
3768 ac
->avail
-= batchcount
;
3769 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3773 * Release an obj back to its cache. If the obj has a constructed state, it must
3774 * be in this state _before_ it is released. Called with disabled ints.
3776 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3779 struct array_cache
*ac
= cpu_cache_get(cachep
);
3782 kmemleak_free_recursive(objp
, cachep
->flags
);
3783 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3785 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3788 * Skip calling cache_free_alien() when the platform is not numa.
3789 * This will avoid cache misses that happen while accessing slabp (which
3790 * is per page memory reference) to get nodeid. Instead use a global
3791 * variable to skip the call, which is mostly likely to be present in
3794 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3797 if (likely(ac
->avail
< ac
->limit
)) {
3798 STATS_INC_FREEHIT(cachep
);
3800 STATS_INC_FREEMISS(cachep
);
3801 cache_flusharray(cachep
, ac
);
3804 ac_put_obj(cachep
, ac
, objp
);
3808 * kmem_cache_alloc - Allocate an object
3809 * @cachep: The cache to allocate from.
3810 * @flags: See kmalloc().
3812 * Allocate an object from this cache. The flags are only relevant
3813 * if the cache has no available objects.
3815 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3817 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3819 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3820 cachep
->object_size
, cachep
->size
, flags
);
3824 EXPORT_SYMBOL(kmem_cache_alloc
);
3826 #ifdef CONFIG_TRACING
3828 kmem_cache_alloc_trace(size_t size
, struct kmem_cache
*cachep
, gfp_t flags
)
3832 ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3834 trace_kmalloc(_RET_IP_
, ret
,
3835 size
, slab_buffer_size(cachep
), flags
);
3838 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3842 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3844 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3845 __builtin_return_address(0));
3847 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3848 cachep
->object_size
, cachep
->size
,
3853 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3855 #ifdef CONFIG_TRACING
3856 void *kmem_cache_alloc_node_trace(size_t size
,
3857 struct kmem_cache
*cachep
,
3863 ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3864 __builtin_return_address(0));
3865 trace_kmalloc_node(_RET_IP_
, ret
,
3866 size
, slab_buffer_size(cachep
),
3870 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3873 static __always_inline
void *
3874 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3876 struct kmem_cache
*cachep
;
3878 cachep
= kmem_find_general_cachep(size
, flags
);
3879 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3881 return kmem_cache_alloc_node_trace(size
, cachep
, flags
, node
);
3884 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3885 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3887 return __do_kmalloc_node(size
, flags
, node
,
3888 __builtin_return_address(0));
3890 EXPORT_SYMBOL(__kmalloc_node
);
3892 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3893 int node
, unsigned long caller
)
3895 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3897 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3899 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3901 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3903 EXPORT_SYMBOL(__kmalloc_node
);
3904 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3905 #endif /* CONFIG_NUMA */
3908 * __do_kmalloc - allocate memory
3909 * @size: how many bytes of memory are required.
3910 * @flags: the type of memory to allocate (see kmalloc).
3911 * @caller: function caller for debug tracking of the caller
3913 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3916 struct kmem_cache
*cachep
;
3919 /* If you want to save a few bytes .text space: replace
3921 * Then kmalloc uses the uninlined functions instead of the inline
3924 cachep
= __find_general_cachep(size
, flags
);
3925 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3927 ret
= __cache_alloc(cachep
, flags
, caller
);
3929 trace_kmalloc((unsigned long) caller
, ret
,
3930 size
, cachep
->size
, flags
);
3936 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3937 void *__kmalloc(size_t size
, gfp_t flags
)
3939 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3941 EXPORT_SYMBOL(__kmalloc
);
3943 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3945 return __do_kmalloc(size
, flags
, (void *)caller
);
3947 EXPORT_SYMBOL(__kmalloc_track_caller
);
3950 void *__kmalloc(size_t size
, gfp_t flags
)
3952 return __do_kmalloc(size
, flags
, NULL
);
3954 EXPORT_SYMBOL(__kmalloc
);
3958 * kmem_cache_free - Deallocate an object
3959 * @cachep: The cache the allocation was from.
3960 * @objp: The previously allocated object.
3962 * Free an object which was previously allocated from this
3965 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3967 unsigned long flags
;
3969 local_irq_save(flags
);
3970 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3971 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3972 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3973 __cache_free(cachep
, objp
, __builtin_return_address(0));
3974 local_irq_restore(flags
);
3976 trace_kmem_cache_free(_RET_IP_
, objp
);
3978 EXPORT_SYMBOL(kmem_cache_free
);
3981 * kfree - free previously allocated memory
3982 * @objp: pointer returned by kmalloc.
3984 * If @objp is NULL, no operation is performed.
3986 * Don't free memory not originally allocated by kmalloc()
3987 * or you will run into trouble.
3989 void kfree(const void *objp
)
3991 struct kmem_cache
*c
;
3992 unsigned long flags
;
3994 trace_kfree(_RET_IP_
, objp
);
3996 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3998 local_irq_save(flags
);
3999 kfree_debugcheck(objp
);
4000 c
= virt_to_cache(objp
);
4001 debug_check_no_locks_freed(objp
, c
->object_size
);
4003 debug_check_no_obj_freed(objp
, c
->object_size
);
4004 __cache_free(c
, (void *)objp
, __builtin_return_address(0));
4005 local_irq_restore(flags
);
4007 EXPORT_SYMBOL(kfree
);
4009 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
4011 return cachep
->object_size
;
4013 EXPORT_SYMBOL(kmem_cache_size
);
4016 * This initializes kmem_list3 or resizes various caches for all nodes.
4018 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
4021 struct kmem_list3
*l3
;
4022 struct array_cache
*new_shared
;
4023 struct array_cache
**new_alien
= NULL
;
4025 for_each_online_node(node
) {
4027 if (use_alien_caches
) {
4028 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
4034 if (cachep
->shared
) {
4035 new_shared
= alloc_arraycache(node
,
4036 cachep
->shared
*cachep
->batchcount
,
4039 free_alien_cache(new_alien
);
4044 l3
= cachep
->nodelists
[node
];
4046 struct array_cache
*shared
= l3
->shared
;
4048 spin_lock_irq(&l3
->list_lock
);
4051 free_block(cachep
, shared
->entry
,
4052 shared
->avail
, node
);
4054 l3
->shared
= new_shared
;
4056 l3
->alien
= new_alien
;
4059 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4060 cachep
->batchcount
+ cachep
->num
;
4061 spin_unlock_irq(&l3
->list_lock
);
4063 free_alien_cache(new_alien
);
4066 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
4068 free_alien_cache(new_alien
);
4073 kmem_list3_init(l3
);
4074 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
4075 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
4076 l3
->shared
= new_shared
;
4077 l3
->alien
= new_alien
;
4078 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4079 cachep
->batchcount
+ cachep
->num
;
4080 cachep
->nodelists
[node
] = l3
;
4085 if (!cachep
->list
.next
) {
4086 /* Cache is not active yet. Roll back what we did */
4089 if (cachep
->nodelists
[node
]) {
4090 l3
= cachep
->nodelists
[node
];
4093 free_alien_cache(l3
->alien
);
4095 cachep
->nodelists
[node
] = NULL
;
4103 struct ccupdate_struct
{
4104 struct kmem_cache
*cachep
;
4105 struct array_cache
*new[0];
4108 static void do_ccupdate_local(void *info
)
4110 struct ccupdate_struct
*new = info
;
4111 struct array_cache
*old
;
4114 old
= cpu_cache_get(new->cachep
);
4116 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
4117 new->new[smp_processor_id()] = old
;
4120 /* Always called with the slab_mutex held */
4121 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
4122 int batchcount
, int shared
, gfp_t gfp
)
4124 struct ccupdate_struct
*new;
4127 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
4132 for_each_online_cpu(i
) {
4133 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
4136 for (i
--; i
>= 0; i
--)
4142 new->cachep
= cachep
;
4144 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
4147 cachep
->batchcount
= batchcount
;
4148 cachep
->limit
= limit
;
4149 cachep
->shared
= shared
;
4151 for_each_online_cpu(i
) {
4152 struct array_cache
*ccold
= new->new[i
];
4155 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4156 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4157 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4161 return alloc_kmemlist(cachep
, gfp
);
4164 /* Called with slab_mutex held always */
4165 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4171 * The head array serves three purposes:
4172 * - create a LIFO ordering, i.e. return objects that are cache-warm
4173 * - reduce the number of spinlock operations.
4174 * - reduce the number of linked list operations on the slab and
4175 * bufctl chains: array operations are cheaper.
4176 * The numbers are guessed, we should auto-tune as described by
4179 if (cachep
->size
> 131072)
4181 else if (cachep
->size
> PAGE_SIZE
)
4183 else if (cachep
->size
> 1024)
4185 else if (cachep
->size
> 256)
4191 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4192 * allocation behaviour: Most allocs on one cpu, most free operations
4193 * on another cpu. For these cases, an efficient object passing between
4194 * cpus is necessary. This is provided by a shared array. The array
4195 * replaces Bonwick's magazine layer.
4196 * On uniprocessor, it's functionally equivalent (but less efficient)
4197 * to a larger limit. Thus disabled by default.
4200 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4205 * With debugging enabled, large batchcount lead to excessively long
4206 * periods with disabled local interrupts. Limit the batchcount
4211 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4213 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4214 cachep
->name
, -err
);
4219 * Drain an array if it contains any elements taking the l3 lock only if
4220 * necessary. Note that the l3 listlock also protects the array_cache
4221 * if drain_array() is used on the shared array.
4223 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4224 struct array_cache
*ac
, int force
, int node
)
4228 if (!ac
|| !ac
->avail
)
4230 if (ac
->touched
&& !force
) {
4233 spin_lock_irq(&l3
->list_lock
);
4235 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4236 if (tofree
> ac
->avail
)
4237 tofree
= (ac
->avail
+ 1) / 2;
4238 free_block(cachep
, ac
->entry
, tofree
, node
);
4239 ac
->avail
-= tofree
;
4240 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4241 sizeof(void *) * ac
->avail
);
4243 spin_unlock_irq(&l3
->list_lock
);
4248 * cache_reap - Reclaim memory from caches.
4249 * @w: work descriptor
4251 * Called from workqueue/eventd every few seconds.
4253 * - clear the per-cpu caches for this CPU.
4254 * - return freeable pages to the main free memory pool.
4256 * If we cannot acquire the cache chain mutex then just give up - we'll try
4257 * again on the next iteration.
4259 static void cache_reap(struct work_struct
*w
)
4261 struct kmem_cache
*searchp
;
4262 struct kmem_list3
*l3
;
4263 int node
= numa_mem_id();
4264 struct delayed_work
*work
= to_delayed_work(w
);
4266 if (!mutex_trylock(&slab_mutex
))
4267 /* Give up. Setup the next iteration. */
4270 list_for_each_entry(searchp
, &slab_caches
, list
) {
4274 * We only take the l3 lock if absolutely necessary and we
4275 * have established with reasonable certainty that
4276 * we can do some work if the lock was obtained.
4278 l3
= searchp
->nodelists
[node
];
4280 reap_alien(searchp
, l3
);
4282 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4285 * These are racy checks but it does not matter
4286 * if we skip one check or scan twice.
4288 if (time_after(l3
->next_reap
, jiffies
))
4291 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4293 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4295 if (l3
->free_touched
)
4296 l3
->free_touched
= 0;
4300 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4301 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4302 STATS_ADD_REAPED(searchp
, freed
);
4308 mutex_unlock(&slab_mutex
);
4311 /* Set up the next iteration */
4312 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4315 #ifdef CONFIG_SLABINFO
4317 static void print_slabinfo_header(struct seq_file
*m
)
4320 * Output format version, so at least we can change it
4321 * without _too_ many complaints.
4324 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4326 seq_puts(m
, "slabinfo - version: 2.1\n");
4328 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4329 "<objperslab> <pagesperslab>");
4330 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4331 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4333 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4334 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4335 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4340 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4344 mutex_lock(&slab_mutex
);
4346 print_slabinfo_header(m
);
4348 return seq_list_start(&slab_caches
, *pos
);
4351 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4353 return seq_list_next(p
, &slab_caches
, pos
);
4356 static void s_stop(struct seq_file
*m
, void *p
)
4358 mutex_unlock(&slab_mutex
);
4361 static int s_show(struct seq_file
*m
, void *p
)
4363 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4365 unsigned long active_objs
;
4366 unsigned long num_objs
;
4367 unsigned long active_slabs
= 0;
4368 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4372 struct kmem_list3
*l3
;
4376 for_each_online_node(node
) {
4377 l3
= cachep
->nodelists
[node
];
4382 spin_lock_irq(&l3
->list_lock
);
4384 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4385 if (slabp
->inuse
!= cachep
->num
&& !error
)
4386 error
= "slabs_full accounting error";
4387 active_objs
+= cachep
->num
;
4390 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4391 if (slabp
->inuse
== cachep
->num
&& !error
)
4392 error
= "slabs_partial inuse accounting error";
4393 if (!slabp
->inuse
&& !error
)
4394 error
= "slabs_partial/inuse accounting error";
4395 active_objs
+= slabp
->inuse
;
4398 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4399 if (slabp
->inuse
&& !error
)
4400 error
= "slabs_free/inuse accounting error";
4403 free_objects
+= l3
->free_objects
;
4405 shared_avail
+= l3
->shared
->avail
;
4407 spin_unlock_irq(&l3
->list_lock
);
4409 num_slabs
+= active_slabs
;
4410 num_objs
= num_slabs
* cachep
->num
;
4411 if (num_objs
- active_objs
!= free_objects
&& !error
)
4412 error
= "free_objects accounting error";
4414 name
= cachep
->name
;
4416 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4418 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4419 name
, active_objs
, num_objs
, cachep
->size
,
4420 cachep
->num
, (1 << cachep
->gfporder
));
4421 seq_printf(m
, " : tunables %4u %4u %4u",
4422 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4423 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4424 active_slabs
, num_slabs
, shared_avail
);
4427 unsigned long high
= cachep
->high_mark
;
4428 unsigned long allocs
= cachep
->num_allocations
;
4429 unsigned long grown
= cachep
->grown
;
4430 unsigned long reaped
= cachep
->reaped
;
4431 unsigned long errors
= cachep
->errors
;
4432 unsigned long max_freeable
= cachep
->max_freeable
;
4433 unsigned long node_allocs
= cachep
->node_allocs
;
4434 unsigned long node_frees
= cachep
->node_frees
;
4435 unsigned long overflows
= cachep
->node_overflow
;
4437 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4438 "%4lu %4lu %4lu %4lu %4lu",
4439 allocs
, high
, grown
,
4440 reaped
, errors
, max_freeable
, node_allocs
,
4441 node_frees
, overflows
);
4445 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4446 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4447 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4448 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4450 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4451 allochit
, allocmiss
, freehit
, freemiss
);
4459 * slabinfo_op - iterator that generates /proc/slabinfo
4468 * num-pages-per-slab
4469 * + further values on SMP and with statistics enabled
4472 static const struct seq_operations slabinfo_op
= {
4479 #define MAX_SLABINFO_WRITE 128
4481 * slabinfo_write - Tuning for the slab allocator
4483 * @buffer: user buffer
4484 * @count: data length
4487 static ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4488 size_t count
, loff_t
*ppos
)
4490 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4491 int limit
, batchcount
, shared
, res
;
4492 struct kmem_cache
*cachep
;
4494 if (count
> MAX_SLABINFO_WRITE
)
4496 if (copy_from_user(&kbuf
, buffer
, count
))
4498 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4500 tmp
= strchr(kbuf
, ' ');
4505 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4508 /* Find the cache in the chain of caches. */
4509 mutex_lock(&slab_mutex
);
4511 list_for_each_entry(cachep
, &slab_caches
, list
) {
4512 if (!strcmp(cachep
->name
, kbuf
)) {
4513 if (limit
< 1 || batchcount
< 1 ||
4514 batchcount
> limit
|| shared
< 0) {
4517 res
= do_tune_cpucache(cachep
, limit
,
4524 mutex_unlock(&slab_mutex
);
4530 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4532 return seq_open(file
, &slabinfo_op
);
4535 static const struct file_operations proc_slabinfo_operations
= {
4536 .open
= slabinfo_open
,
4538 .write
= slabinfo_write
,
4539 .llseek
= seq_lseek
,
4540 .release
= seq_release
,
4543 #ifdef CONFIG_DEBUG_SLAB_LEAK
4545 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4547 mutex_lock(&slab_mutex
);
4548 return seq_list_start(&slab_caches
, *pos
);
4551 static inline int add_caller(unsigned long *n
, unsigned long v
)
4561 unsigned long *q
= p
+ 2 * i
;
4575 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4581 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4587 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4588 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4590 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4595 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4597 #ifdef CONFIG_KALLSYMS
4598 unsigned long offset
, size
;
4599 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4601 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4602 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4604 seq_printf(m
, " [%s]", modname
);
4608 seq_printf(m
, "%p", (void *)address
);
4611 static int leaks_show(struct seq_file
*m
, void *p
)
4613 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4615 struct kmem_list3
*l3
;
4617 unsigned long *n
= m
->private;
4621 if (!(cachep
->flags
& SLAB_STORE_USER
))
4623 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4626 /* OK, we can do it */
4630 for_each_online_node(node
) {
4631 l3
= cachep
->nodelists
[node
];
4636 spin_lock_irq(&l3
->list_lock
);
4638 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4639 handle_slab(n
, cachep
, slabp
);
4640 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4641 handle_slab(n
, cachep
, slabp
);
4642 spin_unlock_irq(&l3
->list_lock
);
4644 name
= cachep
->name
;
4646 /* Increase the buffer size */
4647 mutex_unlock(&slab_mutex
);
4648 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4650 /* Too bad, we are really out */
4652 mutex_lock(&slab_mutex
);
4655 *(unsigned long *)m
->private = n
[0] * 2;
4657 mutex_lock(&slab_mutex
);
4658 /* Now make sure this entry will be retried */
4662 for (i
= 0; i
< n
[1]; i
++) {
4663 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4664 show_symbol(m
, n
[2*i
+2]);
4671 static const struct seq_operations slabstats_op
= {
4672 .start
= leaks_start
,
4678 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4680 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4683 ret
= seq_open(file
, &slabstats_op
);
4685 struct seq_file
*m
= file
->private_data
;
4686 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4695 static const struct file_operations proc_slabstats_operations
= {
4696 .open
= slabstats_open
,
4698 .llseek
= seq_lseek
,
4699 .release
= seq_release_private
,
4703 static int __init
slab_proc_init(void)
4705 proc_create("slabinfo",S_IWUSR
|S_IRUSR
,NULL
,&proc_slabinfo_operations
);
4706 #ifdef CONFIG_DEBUG_SLAB_LEAK
4707 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4711 module_init(slab_proc_init
);
4715 * ksize - get the actual amount of memory allocated for a given object
4716 * @objp: Pointer to the object
4718 * kmalloc may internally round up allocations and return more memory
4719 * than requested. ksize() can be used to determine the actual amount of
4720 * memory allocated. The caller may use this additional memory, even though
4721 * a smaller amount of memory was initially specified with the kmalloc call.
4722 * The caller must guarantee that objp points to a valid object previously
4723 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4724 * must not be freed during the duration of the call.
4726 size_t ksize(const void *objp
)
4729 if (unlikely(objp
== ZERO_SIZE_PTR
))
4732 return virt_to_cache(objp
)->object_size
;
4734 EXPORT_SYMBOL(ksize
);