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 <net/sock.h>
123 #include <asm/cacheflush.h>
124 #include <asm/tlbflush.h>
125 #include <asm/page.h>
127 #include <trace/events/kmem.h>
129 #include "internal.h"
132 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
133 * 0 for faster, smaller code (especially in the critical paths).
135 * STATS - 1 to collect stats for /proc/slabinfo.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 #ifdef CONFIG_DEBUG_SLAB
144 #define FORCED_DEBUG 1
148 #define FORCED_DEBUG 0
151 /* Shouldn't this be in a header file somewhere? */
152 #define BYTES_PER_WORD sizeof(void *)
153 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155 #ifndef ARCH_KMALLOC_FLAGS
156 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 * true if a page was allocated from pfmemalloc reserves for network-based
163 static bool pfmemalloc_active __read_mostly
;
165 /* Legal flag mask for kmem_cache_create(). */
167 # define CREATE_MASK (SLAB_RED_ZONE | \
168 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
171 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
172 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
173 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
175 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
179 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
185 * Bufctl's are used for linking objs within a slab
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 typedef unsigned int kmem_bufctl_t
;
202 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
204 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
205 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
210 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
211 * arrange for kmem_freepages to be called via RCU. This is useful if
212 * we need to approach a kernel structure obliquely, from its address
213 * obtained without the usual locking. We can lock the structure to
214 * stabilize it and check it's still at the given address, only if we
215 * can be sure that the memory has not been meanwhile reused for some
216 * other kind of object (which our subsystem's lock might corrupt).
218 * rcu_read_lock before reading the address, then rcu_read_unlock after
219 * taking the spinlock within the structure expected at that address.
222 struct rcu_head head
;
223 struct kmem_cache
*cachep
;
230 * Manages the objs in a slab. Placed either at the beginning of mem allocated
231 * for a slab, or allocated from an general cache.
232 * Slabs are chained into three list: fully used, partial, fully free slabs.
237 struct list_head list
;
238 unsigned long colouroff
;
239 void *s_mem
; /* including colour offset */
240 unsigned int inuse
; /* num of objs active in slab */
242 unsigned short nodeid
;
244 struct slab_rcu __slab_cover_slab_rcu
;
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
256 * The limit is stored in the per-cpu structure to reduce the data cache
263 unsigned int batchcount
;
264 unsigned int touched
;
267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
271 * Entries should not be directly dereferenced as
272 * entries belonging to slabs marked pfmemalloc will
273 * have the lower bits set SLAB_OBJ_PFMEMALLOC
277 #define SLAB_OBJ_PFMEMALLOC 1
278 static inline bool is_obj_pfmemalloc(void *objp
)
280 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
283 static inline void set_obj_pfmemalloc(void **objp
)
285 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
289 static inline void clear_obj_pfmemalloc(void **objp
)
291 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
295 * bootstrap: The caches do not work without cpuarrays anymore, but the
296 * cpuarrays are allocated from the generic caches...
298 #define BOOT_CPUCACHE_ENTRIES 1
299 struct arraycache_init
{
300 struct array_cache cache
;
301 void *entries
[BOOT_CPUCACHE_ENTRIES
];
305 * The slab lists for all objects.
308 struct list_head slabs_partial
; /* partial list first, better asm code */
309 struct list_head slabs_full
;
310 struct list_head slabs_free
;
311 unsigned long free_objects
;
312 unsigned int free_limit
;
313 unsigned int colour_next
; /* Per-node cache coloring */
314 spinlock_t list_lock
;
315 struct array_cache
*shared
; /* shared per node */
316 struct array_cache
**alien
; /* on other nodes */
317 unsigned long next_reap
; /* updated without locking */
318 int free_touched
; /* updated without locking */
322 * Need this for bootstrapping a per node allocator.
324 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
325 static struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
326 #define CACHE_CACHE 0
327 #define SIZE_AC MAX_NUMNODES
328 #define SIZE_L3 (2 * MAX_NUMNODES)
330 static int drain_freelist(struct kmem_cache
*cache
,
331 struct kmem_list3
*l3
, int tofree
);
332 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
334 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
335 static void cache_reap(struct work_struct
*unused
);
338 * This function must be completely optimized away if a constant is passed to
339 * it. Mostly the same as what is in linux/slab.h except it returns an index.
341 static __always_inline
int index_of(const size_t size
)
343 extern void __bad_size(void);
345 if (__builtin_constant_p(size
)) {
353 #include <linux/kmalloc_sizes.h>
361 static int slab_early_init
= 1;
363 #define INDEX_AC index_of(sizeof(struct arraycache_init))
364 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
366 static void kmem_list3_init(struct kmem_list3
*parent
)
368 INIT_LIST_HEAD(&parent
->slabs_full
);
369 INIT_LIST_HEAD(&parent
->slabs_partial
);
370 INIT_LIST_HEAD(&parent
->slabs_free
);
371 parent
->shared
= NULL
;
372 parent
->alien
= NULL
;
373 parent
->colour_next
= 0;
374 spin_lock_init(&parent
->list_lock
);
375 parent
->free_objects
= 0;
376 parent
->free_touched
= 0;
379 #define MAKE_LIST(cachep, listp, slab, nodeid) \
381 INIT_LIST_HEAD(listp); \
382 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
385 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
387 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
388 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
389 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
392 #define CFLGS_OFF_SLAB (0x80000000UL)
393 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
395 #define BATCHREFILL_LIMIT 16
397 * Optimization question: fewer reaps means less probability for unnessary
398 * cpucache drain/refill cycles.
400 * OTOH the cpuarrays can contain lots of objects,
401 * which could lock up otherwise freeable slabs.
403 #define REAPTIMEOUT_CPUC (2*HZ)
404 #define REAPTIMEOUT_LIST3 (4*HZ)
407 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
408 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
409 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
410 #define STATS_INC_GROWN(x) ((x)->grown++)
411 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
412 #define STATS_SET_HIGH(x) \
414 if ((x)->num_active > (x)->high_mark) \
415 (x)->high_mark = (x)->num_active; \
417 #define STATS_INC_ERR(x) ((x)->errors++)
418 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
419 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
420 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
421 #define STATS_SET_FREEABLE(x, i) \
423 if ((x)->max_freeable < i) \
424 (x)->max_freeable = i; \
426 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
427 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
428 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
429 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
431 #define STATS_INC_ACTIVE(x) do { } while (0)
432 #define STATS_DEC_ACTIVE(x) do { } while (0)
433 #define STATS_INC_ALLOCED(x) do { } while (0)
434 #define STATS_INC_GROWN(x) do { } while (0)
435 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
436 #define STATS_SET_HIGH(x) do { } while (0)
437 #define STATS_INC_ERR(x) do { } while (0)
438 #define STATS_INC_NODEALLOCS(x) do { } while (0)
439 #define STATS_INC_NODEFREES(x) do { } while (0)
440 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
441 #define STATS_SET_FREEABLE(x, i) do { } while (0)
442 #define STATS_INC_ALLOCHIT(x) do { } while (0)
443 #define STATS_INC_ALLOCMISS(x) do { } while (0)
444 #define STATS_INC_FREEHIT(x) do { } while (0)
445 #define STATS_INC_FREEMISS(x) do { } while (0)
451 * memory layout of objects:
453 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
454 * the end of an object is aligned with the end of the real
455 * allocation. Catches writes behind the end of the allocation.
456 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
458 * cachep->obj_offset: The real object.
459 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
460 * cachep->size - 1* BYTES_PER_WORD: last caller address
461 * [BYTES_PER_WORD long]
463 static int obj_offset(struct kmem_cache
*cachep
)
465 return cachep
->obj_offset
;
468 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
470 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
471 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
472 sizeof(unsigned long long));
475 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
477 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
478 if (cachep
->flags
& SLAB_STORE_USER
)
479 return (unsigned long long *)(objp
+ cachep
->size
-
480 sizeof(unsigned long long) -
482 return (unsigned long long *) (objp
+ cachep
->size
-
483 sizeof(unsigned long long));
486 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
488 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
489 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
494 #define obj_offset(x) 0
495 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
496 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
497 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
501 #ifdef CONFIG_TRACING
502 size_t slab_buffer_size(struct kmem_cache
*cachep
)
506 EXPORT_SYMBOL(slab_buffer_size
);
510 * Do not go above this order unless 0 objects fit into the slab or
511 * overridden on the command line.
513 #define SLAB_MAX_ORDER_HI 1
514 #define SLAB_MAX_ORDER_LO 0
515 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
516 static bool slab_max_order_set __initdata
;
518 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
520 struct page
*page
= virt_to_head_page(obj
);
521 return page
->slab_cache
;
524 static inline struct slab
*virt_to_slab(const void *obj
)
526 struct page
*page
= virt_to_head_page(obj
);
528 VM_BUG_ON(!PageSlab(page
));
529 return page
->slab_page
;
532 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
535 return slab
->s_mem
+ cache
->size
* idx
;
539 * We want to avoid an expensive divide : (offset / cache->size)
540 * Using the fact that size is a constant for a particular cache,
541 * we can replace (offset / cache->size) by
542 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
544 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
545 const struct slab
*slab
, void *obj
)
547 u32 offset
= (obj
- slab
->s_mem
);
548 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
552 * These are the default caches for kmalloc. Custom caches can have other sizes.
554 struct cache_sizes malloc_sizes
[] = {
555 #define CACHE(x) { .cs_size = (x) },
556 #include <linux/kmalloc_sizes.h>
560 EXPORT_SYMBOL(malloc_sizes
);
562 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
568 static struct cache_names __initdata cache_names
[] = {
569 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
570 #include <linux/kmalloc_sizes.h>
575 static struct arraycache_init initarray_cache __initdata
=
576 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
577 static struct arraycache_init initarray_generic
=
578 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
580 /* internal cache of cache description objs */
581 static struct kmem_list3
*kmem_cache_nodelists
[MAX_NUMNODES
];
582 static struct kmem_cache kmem_cache_boot
= {
583 .nodelists
= kmem_cache_nodelists
,
585 .limit
= BOOT_CPUCACHE_ENTRIES
,
587 .size
= sizeof(struct kmem_cache
),
588 .name
= "kmem_cache",
591 #define BAD_ALIEN_MAGIC 0x01020304ul
593 #ifdef CONFIG_LOCKDEP
596 * Slab sometimes uses the kmalloc slabs to store the slab headers
597 * for other slabs "off slab".
598 * The locking for this is tricky in that it nests within the locks
599 * of all other slabs in a few places; to deal with this special
600 * locking we put on-slab caches into a separate lock-class.
602 * We set lock class for alien array caches which are up during init.
603 * The lock annotation will be lost if all cpus of a node goes down and
604 * then comes back up during hotplug
606 static struct lock_class_key on_slab_l3_key
;
607 static struct lock_class_key on_slab_alc_key
;
609 static struct lock_class_key debugobj_l3_key
;
610 static struct lock_class_key debugobj_alc_key
;
612 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
613 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
616 struct array_cache
**alc
;
617 struct kmem_list3
*l3
;
620 l3
= cachep
->nodelists
[q
];
624 lockdep_set_class(&l3
->list_lock
, l3_key
);
627 * FIXME: This check for BAD_ALIEN_MAGIC
628 * should go away when common slab code is taught to
629 * work even without alien caches.
630 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
631 * for alloc_alien_cache,
633 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
637 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
641 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
643 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
646 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
650 for_each_online_node(node
)
651 slab_set_debugobj_lock_classes_node(cachep
, node
);
654 static void init_node_lock_keys(int q
)
656 struct cache_sizes
*s
= malloc_sizes
;
661 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
662 struct kmem_list3
*l3
;
664 l3
= s
->cs_cachep
->nodelists
[q
];
665 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
668 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
669 &on_slab_alc_key
, q
);
673 static inline void init_lock_keys(void)
678 init_node_lock_keys(node
);
681 static void init_node_lock_keys(int q
)
685 static inline void init_lock_keys(void)
689 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
693 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
698 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
700 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
702 return cachep
->array
[smp_processor_id()];
705 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
708 struct cache_sizes
*csizep
= malloc_sizes
;
711 /* This happens if someone tries to call
712 * kmem_cache_create(), or __kmalloc(), before
713 * the generic caches are initialized.
715 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
718 return ZERO_SIZE_PTR
;
720 while (size
> csizep
->cs_size
)
724 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
725 * has cs_{dma,}cachep==NULL. Thus no special case
726 * for large kmalloc calls required.
728 #ifdef CONFIG_ZONE_DMA
729 if (unlikely(gfpflags
& GFP_DMA
))
730 return csizep
->cs_dmacachep
;
732 return csizep
->cs_cachep
;
735 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
737 return __find_general_cachep(size
, gfpflags
);
740 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
742 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
746 * Calculate the number of objects and left-over bytes for a given buffer size.
748 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
749 size_t align
, int flags
, size_t *left_over
,
754 size_t slab_size
= PAGE_SIZE
<< gfporder
;
757 * The slab management structure can be either off the slab or
758 * on it. For the latter case, the memory allocated for a
762 * - One kmem_bufctl_t for each object
763 * - Padding to respect alignment of @align
764 * - @buffer_size bytes for each object
766 * If the slab management structure is off the slab, then the
767 * alignment will already be calculated into the size. Because
768 * the slabs are all pages aligned, the objects will be at the
769 * correct alignment when allocated.
771 if (flags
& CFLGS_OFF_SLAB
) {
773 nr_objs
= slab_size
/ buffer_size
;
775 if (nr_objs
> SLAB_LIMIT
)
776 nr_objs
= SLAB_LIMIT
;
779 * Ignore padding for the initial guess. The padding
780 * is at most @align-1 bytes, and @buffer_size is at
781 * least @align. In the worst case, this result will
782 * be one greater than the number of objects that fit
783 * into the memory allocation when taking the padding
786 nr_objs
= (slab_size
- sizeof(struct slab
)) /
787 (buffer_size
+ sizeof(kmem_bufctl_t
));
790 * This calculated number will be either the right
791 * amount, or one greater than what we want.
793 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
797 if (nr_objs
> SLAB_LIMIT
)
798 nr_objs
= SLAB_LIMIT
;
800 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
803 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
806 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
808 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
811 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
812 function
, cachep
->name
, msg
);
817 * By default on NUMA we use alien caches to stage the freeing of
818 * objects allocated from other nodes. This causes massive memory
819 * inefficiencies when using fake NUMA setup to split memory into a
820 * large number of small nodes, so it can be disabled on the command
824 static int use_alien_caches __read_mostly
= 1;
825 static int __init
noaliencache_setup(char *s
)
827 use_alien_caches
= 0;
830 __setup("noaliencache", noaliencache_setup
);
832 static int __init
slab_max_order_setup(char *str
)
834 get_option(&str
, &slab_max_order
);
835 slab_max_order
= slab_max_order
< 0 ? 0 :
836 min(slab_max_order
, MAX_ORDER
- 1);
837 slab_max_order_set
= true;
841 __setup("slab_max_order=", slab_max_order_setup
);
845 * Special reaping functions for NUMA systems called from cache_reap().
846 * These take care of doing round robin flushing of alien caches (containing
847 * objects freed on different nodes from which they were allocated) and the
848 * flushing of remote pcps by calling drain_node_pages.
850 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
852 static void init_reap_node(int cpu
)
856 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
857 if (node
== MAX_NUMNODES
)
858 node
= first_node(node_online_map
);
860 per_cpu(slab_reap_node
, cpu
) = node
;
863 static void next_reap_node(void)
865 int node
= __this_cpu_read(slab_reap_node
);
867 node
= next_node(node
, node_online_map
);
868 if (unlikely(node
>= MAX_NUMNODES
))
869 node
= first_node(node_online_map
);
870 __this_cpu_write(slab_reap_node
, node
);
874 #define init_reap_node(cpu) do { } while (0)
875 #define next_reap_node(void) do { } while (0)
879 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
880 * via the workqueue/eventd.
881 * Add the CPU number into the expiration time to minimize the possibility of
882 * the CPUs getting into lockstep and contending for the global cache chain
885 static void __cpuinit
start_cpu_timer(int cpu
)
887 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
890 * When this gets called from do_initcalls via cpucache_init(),
891 * init_workqueues() has already run, so keventd will be setup
894 if (keventd_up() && reap_work
->work
.func
== NULL
) {
896 INIT_DELAYED_WORK_DEFERRABLE(reap_work
, cache_reap
);
897 schedule_delayed_work_on(cpu
, reap_work
,
898 __round_jiffies_relative(HZ
, cpu
));
902 static struct array_cache
*alloc_arraycache(int node
, int entries
,
903 int batchcount
, gfp_t gfp
)
905 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
906 struct array_cache
*nc
= NULL
;
908 nc
= kmalloc_node(memsize
, gfp
, node
);
910 * The array_cache structures contain pointers to free object.
911 * However, when such objects are allocated or transferred to another
912 * cache the pointers are not cleared and they could be counted as
913 * valid references during a kmemleak scan. Therefore, kmemleak must
914 * not scan such objects.
916 kmemleak_no_scan(nc
);
920 nc
->batchcount
= batchcount
;
922 spin_lock_init(&nc
->lock
);
927 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
929 struct page
*page
= virt_to_page(slabp
->s_mem
);
931 return PageSlabPfmemalloc(page
);
934 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
935 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
936 struct array_cache
*ac
)
938 struct kmem_list3
*l3
= cachep
->nodelists
[numa_mem_id()];
942 if (!pfmemalloc_active
)
945 spin_lock_irqsave(&l3
->list_lock
, flags
);
946 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
947 if (is_slab_pfmemalloc(slabp
))
950 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
951 if (is_slab_pfmemalloc(slabp
))
954 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
955 if (is_slab_pfmemalloc(slabp
))
958 pfmemalloc_active
= false;
960 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
963 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
964 gfp_t flags
, bool force_refill
)
967 void *objp
= ac
->entry
[--ac
->avail
];
969 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
970 if (unlikely(is_obj_pfmemalloc(objp
))) {
971 struct kmem_list3
*l3
;
973 if (gfp_pfmemalloc_allowed(flags
)) {
974 clear_obj_pfmemalloc(&objp
);
978 /* The caller cannot use PFMEMALLOC objects, find another one */
979 for (i
= 1; i
< ac
->avail
; i
++) {
980 /* If a !PFMEMALLOC object is found, swap them */
981 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
983 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
984 ac
->entry
[ac
->avail
] = objp
;
990 * If there are empty slabs on the slabs_free list and we are
991 * being forced to refill the cache, mark this one !pfmemalloc.
993 l3
= cachep
->nodelists
[numa_mem_id()];
994 if (!list_empty(&l3
->slabs_free
) && force_refill
) {
995 struct slab
*slabp
= virt_to_slab(objp
);
996 ClearPageSlabPfmemalloc(virt_to_page(slabp
->s_mem
));
997 clear_obj_pfmemalloc(&objp
);
998 recheck_pfmemalloc_active(cachep
, ac
);
1002 /* No !PFMEMALLOC objects available */
1010 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
1011 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
1015 if (unlikely(sk_memalloc_socks()))
1016 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
1018 objp
= ac
->entry
[--ac
->avail
];
1023 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1026 if (unlikely(pfmemalloc_active
)) {
1027 /* Some pfmemalloc slabs exist, check if this is one */
1028 struct page
*page
= virt_to_page(objp
);
1029 if (PageSlabPfmemalloc(page
))
1030 set_obj_pfmemalloc(&objp
);
1036 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1039 if (unlikely(sk_memalloc_socks()))
1040 objp
= __ac_put_obj(cachep
, ac
, objp
);
1042 ac
->entry
[ac
->avail
++] = objp
;
1046 * Transfer objects in one arraycache to another.
1047 * Locking must be handled by the caller.
1049 * Return the number of entries transferred.
1051 static int transfer_objects(struct array_cache
*to
,
1052 struct array_cache
*from
, unsigned int max
)
1054 /* Figure out how many entries to transfer */
1055 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
1060 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1061 sizeof(void *) *nr
);
1070 #define drain_alien_cache(cachep, alien) do { } while (0)
1071 #define reap_alien(cachep, l3) do { } while (0)
1073 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1075 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1078 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1082 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1087 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1093 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1094 gfp_t flags
, int nodeid
)
1099 #else /* CONFIG_NUMA */
1101 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1102 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1104 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1106 struct array_cache
**ac_ptr
;
1107 int memsize
= sizeof(void *) * nr_node_ids
;
1112 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1115 if (i
== node
|| !node_online(i
))
1117 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1119 for (i
--; i
>= 0; i
--)
1129 static void free_alien_cache(struct array_cache
**ac_ptr
)
1140 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1141 struct array_cache
*ac
, int node
)
1143 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1146 spin_lock(&rl3
->list_lock
);
1148 * Stuff objects into the remote nodes shared array first.
1149 * That way we could avoid the overhead of putting the objects
1150 * into the free lists and getting them back later.
1153 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1155 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1157 spin_unlock(&rl3
->list_lock
);
1162 * Called from cache_reap() to regularly drain alien caches round robin.
1164 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1166 int node
= __this_cpu_read(slab_reap_node
);
1169 struct array_cache
*ac
= l3
->alien
[node
];
1171 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1172 __drain_alien_cache(cachep
, ac
, node
);
1173 spin_unlock_irq(&ac
->lock
);
1178 static void drain_alien_cache(struct kmem_cache
*cachep
,
1179 struct array_cache
**alien
)
1182 struct array_cache
*ac
;
1183 unsigned long flags
;
1185 for_each_online_node(i
) {
1188 spin_lock_irqsave(&ac
->lock
, flags
);
1189 __drain_alien_cache(cachep
, ac
, i
);
1190 spin_unlock_irqrestore(&ac
->lock
, flags
);
1195 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1197 struct slab
*slabp
= virt_to_slab(objp
);
1198 int nodeid
= slabp
->nodeid
;
1199 struct kmem_list3
*l3
;
1200 struct array_cache
*alien
= NULL
;
1203 node
= numa_mem_id();
1206 * Make sure we are not freeing a object from another node to the array
1207 * cache on this cpu.
1209 if (likely(slabp
->nodeid
== node
))
1212 l3
= cachep
->nodelists
[node
];
1213 STATS_INC_NODEFREES(cachep
);
1214 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1215 alien
= l3
->alien
[nodeid
];
1216 spin_lock(&alien
->lock
);
1217 if (unlikely(alien
->avail
== alien
->limit
)) {
1218 STATS_INC_ACOVERFLOW(cachep
);
1219 __drain_alien_cache(cachep
, alien
, nodeid
);
1221 ac_put_obj(cachep
, alien
, objp
);
1222 spin_unlock(&alien
->lock
);
1224 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1225 free_block(cachep
, &objp
, 1, nodeid
);
1226 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1233 * Allocates and initializes nodelists for a node on each slab cache, used for
1234 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1235 * will be allocated off-node since memory is not yet online for the new node.
1236 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1239 * Must hold slab_mutex.
1241 static int init_cache_nodelists_node(int node
)
1243 struct kmem_cache
*cachep
;
1244 struct kmem_list3
*l3
;
1245 const int memsize
= sizeof(struct kmem_list3
);
1247 list_for_each_entry(cachep
, &slab_caches
, list
) {
1249 * Set up the size64 kmemlist for cpu before we can
1250 * begin anything. Make sure some other cpu on this
1251 * node has not already allocated this
1253 if (!cachep
->nodelists
[node
]) {
1254 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1257 kmem_list3_init(l3
);
1258 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1259 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1262 * The l3s don't come and go as CPUs come and
1263 * go. slab_mutex is sufficient
1266 cachep
->nodelists
[node
] = l3
;
1269 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1270 cachep
->nodelists
[node
]->free_limit
=
1271 (1 + nr_cpus_node(node
)) *
1272 cachep
->batchcount
+ cachep
->num
;
1273 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1278 static void __cpuinit
cpuup_canceled(long cpu
)
1280 struct kmem_cache
*cachep
;
1281 struct kmem_list3
*l3
= NULL
;
1282 int node
= cpu_to_mem(cpu
);
1283 const struct cpumask
*mask
= cpumask_of_node(node
);
1285 list_for_each_entry(cachep
, &slab_caches
, list
) {
1286 struct array_cache
*nc
;
1287 struct array_cache
*shared
;
1288 struct array_cache
**alien
;
1290 /* cpu is dead; no one can alloc from it. */
1291 nc
= cachep
->array
[cpu
];
1292 cachep
->array
[cpu
] = NULL
;
1293 l3
= cachep
->nodelists
[node
];
1296 goto free_array_cache
;
1298 spin_lock_irq(&l3
->list_lock
);
1300 /* Free limit for this kmem_list3 */
1301 l3
->free_limit
-= cachep
->batchcount
;
1303 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1305 if (!cpumask_empty(mask
)) {
1306 spin_unlock_irq(&l3
->list_lock
);
1307 goto free_array_cache
;
1310 shared
= l3
->shared
;
1312 free_block(cachep
, shared
->entry
,
1313 shared
->avail
, node
);
1320 spin_unlock_irq(&l3
->list_lock
);
1324 drain_alien_cache(cachep
, alien
);
1325 free_alien_cache(alien
);
1331 * In the previous loop, all the objects were freed to
1332 * the respective cache's slabs, now we can go ahead and
1333 * shrink each nodelist to its limit.
1335 list_for_each_entry(cachep
, &slab_caches
, list
) {
1336 l3
= cachep
->nodelists
[node
];
1339 drain_freelist(cachep
, l3
, l3
->free_objects
);
1343 static int __cpuinit
cpuup_prepare(long cpu
)
1345 struct kmem_cache
*cachep
;
1346 struct kmem_list3
*l3
= NULL
;
1347 int node
= cpu_to_mem(cpu
);
1351 * We need to do this right in the beginning since
1352 * alloc_arraycache's are going to use this list.
1353 * kmalloc_node allows us to add the slab to the right
1354 * kmem_list3 and not this cpu's kmem_list3
1356 err
= init_cache_nodelists_node(node
);
1361 * Now we can go ahead with allocating the shared arrays and
1364 list_for_each_entry(cachep
, &slab_caches
, list
) {
1365 struct array_cache
*nc
;
1366 struct array_cache
*shared
= NULL
;
1367 struct array_cache
**alien
= NULL
;
1369 nc
= alloc_arraycache(node
, cachep
->limit
,
1370 cachep
->batchcount
, GFP_KERNEL
);
1373 if (cachep
->shared
) {
1374 shared
= alloc_arraycache(node
,
1375 cachep
->shared
* cachep
->batchcount
,
1376 0xbaadf00d, GFP_KERNEL
);
1382 if (use_alien_caches
) {
1383 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1390 cachep
->array
[cpu
] = nc
;
1391 l3
= cachep
->nodelists
[node
];
1394 spin_lock_irq(&l3
->list_lock
);
1397 * We are serialised from CPU_DEAD or
1398 * CPU_UP_CANCELLED by the cpucontrol lock
1400 l3
->shared
= shared
;
1409 spin_unlock_irq(&l3
->list_lock
);
1411 free_alien_cache(alien
);
1412 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1413 slab_set_debugobj_lock_classes_node(cachep
, node
);
1415 init_node_lock_keys(node
);
1419 cpuup_canceled(cpu
);
1423 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1424 unsigned long action
, void *hcpu
)
1426 long cpu
= (long)hcpu
;
1430 case CPU_UP_PREPARE
:
1431 case CPU_UP_PREPARE_FROZEN
:
1432 mutex_lock(&slab_mutex
);
1433 err
= cpuup_prepare(cpu
);
1434 mutex_unlock(&slab_mutex
);
1437 case CPU_ONLINE_FROZEN
:
1438 start_cpu_timer(cpu
);
1440 #ifdef CONFIG_HOTPLUG_CPU
1441 case CPU_DOWN_PREPARE
:
1442 case CPU_DOWN_PREPARE_FROZEN
:
1444 * Shutdown cache reaper. Note that the slab_mutex is
1445 * held so that if cache_reap() is invoked it cannot do
1446 * anything expensive but will only modify reap_work
1447 * and reschedule the timer.
1449 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1450 /* Now the cache_reaper is guaranteed to be not running. */
1451 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1453 case CPU_DOWN_FAILED
:
1454 case CPU_DOWN_FAILED_FROZEN
:
1455 start_cpu_timer(cpu
);
1458 case CPU_DEAD_FROZEN
:
1460 * Even if all the cpus of a node are down, we don't free the
1461 * kmem_list3 of any cache. This to avoid a race between
1462 * cpu_down, and a kmalloc allocation from another cpu for
1463 * memory from the node of the cpu going down. The list3
1464 * structure is usually allocated from kmem_cache_create() and
1465 * gets destroyed at kmem_cache_destroy().
1469 case CPU_UP_CANCELED
:
1470 case CPU_UP_CANCELED_FROZEN
:
1471 mutex_lock(&slab_mutex
);
1472 cpuup_canceled(cpu
);
1473 mutex_unlock(&slab_mutex
);
1476 return notifier_from_errno(err
);
1479 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1480 &cpuup_callback
, NULL
, 0
1483 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1485 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1486 * Returns -EBUSY if all objects cannot be drained so that the node is not
1489 * Must hold slab_mutex.
1491 static int __meminit
drain_cache_nodelists_node(int node
)
1493 struct kmem_cache
*cachep
;
1496 list_for_each_entry(cachep
, &slab_caches
, list
) {
1497 struct kmem_list3
*l3
;
1499 l3
= cachep
->nodelists
[node
];
1503 drain_freelist(cachep
, l3
, l3
->free_objects
);
1505 if (!list_empty(&l3
->slabs_full
) ||
1506 !list_empty(&l3
->slabs_partial
)) {
1514 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1515 unsigned long action
, void *arg
)
1517 struct memory_notify
*mnb
= arg
;
1521 nid
= mnb
->status_change_nid
;
1526 case MEM_GOING_ONLINE
:
1527 mutex_lock(&slab_mutex
);
1528 ret
= init_cache_nodelists_node(nid
);
1529 mutex_unlock(&slab_mutex
);
1531 case MEM_GOING_OFFLINE
:
1532 mutex_lock(&slab_mutex
);
1533 ret
= drain_cache_nodelists_node(nid
);
1534 mutex_unlock(&slab_mutex
);
1538 case MEM_CANCEL_ONLINE
:
1539 case MEM_CANCEL_OFFLINE
:
1543 return notifier_from_errno(ret
);
1545 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1548 * swap the static kmem_list3 with kmalloced memory
1550 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1553 struct kmem_list3
*ptr
;
1555 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1558 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1560 * Do not assume that spinlocks can be initialized via memcpy:
1562 spin_lock_init(&ptr
->list_lock
);
1564 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1565 cachep
->nodelists
[nodeid
] = ptr
;
1569 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1570 * size of kmem_list3.
1572 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1576 for_each_online_node(node
) {
1577 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1578 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1580 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1585 * Initialisation. Called after the page allocator have been initialised and
1586 * before smp_init().
1588 void __init
kmem_cache_init(void)
1591 struct cache_sizes
*sizes
;
1592 struct cache_names
*names
;
1597 kmem_cache
= &kmem_cache_boot
;
1599 if (num_possible_nodes() == 1)
1600 use_alien_caches
= 0;
1602 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1603 kmem_list3_init(&initkmem_list3
[i
]);
1604 if (i
< MAX_NUMNODES
)
1605 kmem_cache
->nodelists
[i
] = NULL
;
1607 set_up_list3s(kmem_cache
, CACHE_CACHE
);
1610 * Fragmentation resistance on low memory - only use bigger
1611 * page orders on machines with more than 32MB of memory if
1612 * not overridden on the command line.
1614 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1615 slab_max_order
= SLAB_MAX_ORDER_HI
;
1617 /* Bootstrap is tricky, because several objects are allocated
1618 * from caches that do not exist yet:
1619 * 1) initialize the kmem_cache cache: it contains the struct
1620 * kmem_cache structures of all caches, except kmem_cache itself:
1621 * kmem_cache is statically allocated.
1622 * Initially an __init data area is used for the head array and the
1623 * kmem_list3 structures, it's replaced with a kmalloc allocated
1624 * array at the end of the bootstrap.
1625 * 2) Create the first kmalloc cache.
1626 * The struct kmem_cache for the new cache is allocated normally.
1627 * An __init data area is used for the head array.
1628 * 3) Create the remaining kmalloc caches, with minimally sized
1630 * 4) Replace the __init data head arrays for kmem_cache and the first
1631 * kmalloc cache with kmalloc allocated arrays.
1632 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1633 * the other cache's with kmalloc allocated memory.
1634 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1637 node
= numa_mem_id();
1639 /* 1) create the kmem_cache */
1640 INIT_LIST_HEAD(&slab_caches
);
1641 list_add(&kmem_cache
->list
, &slab_caches
);
1642 kmem_cache
->colour_off
= cache_line_size();
1643 kmem_cache
->array
[smp_processor_id()] = &initarray_cache
.cache
;
1644 kmem_cache
->nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1647 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1649 kmem_cache
->size
= offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1650 nr_node_ids
* sizeof(struct kmem_list3
*);
1651 kmem_cache
->object_size
= kmem_cache
->size
;
1652 kmem_cache
->size
= ALIGN(kmem_cache
->object_size
,
1654 kmem_cache
->reciprocal_buffer_size
=
1655 reciprocal_value(kmem_cache
->size
);
1657 for (order
= 0; order
< MAX_ORDER
; order
++) {
1658 cache_estimate(order
, kmem_cache
->size
,
1659 cache_line_size(), 0, &left_over
, &kmem_cache
->num
);
1660 if (kmem_cache
->num
)
1663 BUG_ON(!kmem_cache
->num
);
1664 kmem_cache
->gfporder
= order
;
1665 kmem_cache
->colour
= left_over
/ kmem_cache
->colour_off
;
1666 kmem_cache
->slab_size
= ALIGN(kmem_cache
->num
* sizeof(kmem_bufctl_t
) +
1667 sizeof(struct slab
), cache_line_size());
1669 /* 2+3) create the kmalloc caches */
1670 sizes
= malloc_sizes
;
1671 names
= cache_names
;
1674 * Initialize the caches that provide memory for the array cache and the
1675 * kmem_list3 structures first. Without this, further allocations will
1679 sizes
[INDEX_AC
].cs_cachep
= __kmem_cache_create(names
[INDEX_AC
].name
,
1680 sizes
[INDEX_AC
].cs_size
,
1681 ARCH_KMALLOC_MINALIGN
,
1682 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1685 list_add(&sizes
[INDEX_AC
].cs_cachep
->list
, &slab_caches
);
1686 if (INDEX_AC
!= INDEX_L3
) {
1687 sizes
[INDEX_L3
].cs_cachep
=
1688 __kmem_cache_create(names
[INDEX_L3
].name
,
1689 sizes
[INDEX_L3
].cs_size
,
1690 ARCH_KMALLOC_MINALIGN
,
1691 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1693 list_add(&sizes
[INDEX_L3
].cs_cachep
->list
, &slab_caches
);
1696 slab_early_init
= 0;
1698 while (sizes
->cs_size
!= ULONG_MAX
) {
1700 * For performance, all the general caches are L1 aligned.
1701 * This should be particularly beneficial on SMP boxes, as it
1702 * eliminates "false sharing".
1703 * Note for systems short on memory removing the alignment will
1704 * allow tighter packing of the smaller caches.
1706 if (!sizes
->cs_cachep
) {
1707 sizes
->cs_cachep
= __kmem_cache_create(names
->name
,
1709 ARCH_KMALLOC_MINALIGN
,
1710 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1712 list_add(&sizes
->cs_cachep
->list
, &slab_caches
);
1714 #ifdef CONFIG_ZONE_DMA
1715 sizes
->cs_dmacachep
= __kmem_cache_create(
1718 ARCH_KMALLOC_MINALIGN
,
1719 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1722 list_add(&sizes
->cs_dmacachep
->list
, &slab_caches
);
1727 /* 4) Replace the bootstrap head arrays */
1729 struct array_cache
*ptr
;
1731 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1733 BUG_ON(cpu_cache_get(kmem_cache
) != &initarray_cache
.cache
);
1734 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1735 sizeof(struct arraycache_init
));
1737 * Do not assume that spinlocks can be initialized via memcpy:
1739 spin_lock_init(&ptr
->lock
);
1741 kmem_cache
->array
[smp_processor_id()] = ptr
;
1743 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1745 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1746 != &initarray_generic
.cache
);
1747 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1748 sizeof(struct arraycache_init
));
1750 * Do not assume that spinlocks can be initialized via memcpy:
1752 spin_lock_init(&ptr
->lock
);
1754 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1757 /* 5) Replace the bootstrap kmem_list3's */
1761 for_each_online_node(nid
) {
1762 init_list(kmem_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1764 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1765 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1767 if (INDEX_AC
!= INDEX_L3
) {
1768 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1769 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1777 void __init
kmem_cache_init_late(void)
1779 struct kmem_cache
*cachep
;
1783 /* Annotate slab for lockdep -- annotate the malloc caches */
1786 /* 6) resize the head arrays to their final sizes */
1787 mutex_lock(&slab_mutex
);
1788 list_for_each_entry(cachep
, &slab_caches
, list
)
1789 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1791 mutex_unlock(&slab_mutex
);
1797 * Register a cpu startup notifier callback that initializes
1798 * cpu_cache_get for all new cpus
1800 register_cpu_notifier(&cpucache_notifier
);
1804 * Register a memory hotplug callback that initializes and frees
1807 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1811 * The reap timers are started later, with a module init call: That part
1812 * of the kernel is not yet operational.
1816 static int __init
cpucache_init(void)
1821 * Register the timers that return unneeded pages to the page allocator
1823 for_each_online_cpu(cpu
)
1824 start_cpu_timer(cpu
);
1830 __initcall(cpucache_init
);
1832 static noinline
void
1833 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1835 struct kmem_list3
*l3
;
1837 unsigned long flags
;
1841 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1843 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1844 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1846 for_each_online_node(node
) {
1847 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1848 unsigned long active_slabs
= 0, num_slabs
= 0;
1850 l3
= cachep
->nodelists
[node
];
1854 spin_lock_irqsave(&l3
->list_lock
, flags
);
1855 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
1856 active_objs
+= cachep
->num
;
1859 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
1860 active_objs
+= slabp
->inuse
;
1863 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
1866 free_objects
+= l3
->free_objects
;
1867 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
1869 num_slabs
+= active_slabs
;
1870 num_objs
= num_slabs
* cachep
->num
;
1872 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1873 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1879 * Interface to system's page allocator. No need to hold the cache-lock.
1881 * If we requested dmaable memory, we will get it. Even if we
1882 * did not request dmaable memory, we might get it, but that
1883 * would be relatively rare and ignorable.
1885 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1893 * Nommu uses slab's for process anonymous memory allocations, and thus
1894 * requires __GFP_COMP to properly refcount higher order allocations
1896 flags
|= __GFP_COMP
;
1899 flags
|= cachep
->allocflags
;
1900 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1901 flags
|= __GFP_RECLAIMABLE
;
1903 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1905 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1906 slab_out_of_memory(cachep
, flags
, nodeid
);
1910 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1911 if (unlikely(page
->pfmemalloc
))
1912 pfmemalloc_active
= true;
1914 nr_pages
= (1 << cachep
->gfporder
);
1915 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1916 add_zone_page_state(page_zone(page
),
1917 NR_SLAB_RECLAIMABLE
, nr_pages
);
1919 add_zone_page_state(page_zone(page
),
1920 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1921 for (i
= 0; i
< nr_pages
; i
++) {
1922 __SetPageSlab(page
+ i
);
1924 if (page
->pfmemalloc
)
1925 SetPageSlabPfmemalloc(page
+ i
);
1928 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1929 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1932 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1934 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1937 return page_address(page
);
1941 * Interface to system's page release.
1943 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1945 unsigned long i
= (1 << cachep
->gfporder
);
1946 struct page
*page
= virt_to_page(addr
);
1947 const unsigned long nr_freed
= i
;
1949 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1951 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1952 sub_zone_page_state(page_zone(page
),
1953 NR_SLAB_RECLAIMABLE
, nr_freed
);
1955 sub_zone_page_state(page_zone(page
),
1956 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1958 BUG_ON(!PageSlab(page
));
1959 __ClearPageSlabPfmemalloc(page
);
1960 __ClearPageSlab(page
);
1963 if (current
->reclaim_state
)
1964 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1965 free_pages((unsigned long)addr
, cachep
->gfporder
);
1968 static void kmem_rcu_free(struct rcu_head
*head
)
1970 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1971 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1973 kmem_freepages(cachep
, slab_rcu
->addr
);
1974 if (OFF_SLAB(cachep
))
1975 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1980 #ifdef CONFIG_DEBUG_PAGEALLOC
1981 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1982 unsigned long caller
)
1984 int size
= cachep
->object_size
;
1986 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1988 if (size
< 5 * sizeof(unsigned long))
1991 *addr
++ = 0x12345678;
1993 *addr
++ = smp_processor_id();
1994 size
-= 3 * sizeof(unsigned long);
1996 unsigned long *sptr
= &caller
;
1997 unsigned long svalue
;
1999 while (!kstack_end(sptr
)) {
2001 if (kernel_text_address(svalue
)) {
2003 size
-= sizeof(unsigned long);
2004 if (size
<= sizeof(unsigned long))
2010 *addr
++ = 0x87654321;
2014 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
2016 int size
= cachep
->object_size
;
2017 addr
= &((char *)addr
)[obj_offset(cachep
)];
2019 memset(addr
, val
, size
);
2020 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
2023 static void dump_line(char *data
, int offset
, int limit
)
2026 unsigned char error
= 0;
2029 printk(KERN_ERR
"%03x: ", offset
);
2030 for (i
= 0; i
< limit
; i
++) {
2031 if (data
[offset
+ i
] != POISON_FREE
) {
2032 error
= data
[offset
+ i
];
2036 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
2037 &data
[offset
], limit
, 1);
2039 if (bad_count
== 1) {
2040 error
^= POISON_FREE
;
2041 if (!(error
& (error
- 1))) {
2042 printk(KERN_ERR
"Single bit error detected. Probably "
2045 printk(KERN_ERR
"Run memtest86+ or a similar memory "
2048 printk(KERN_ERR
"Run a memory test tool.\n");
2057 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
2062 if (cachep
->flags
& SLAB_RED_ZONE
) {
2063 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
2064 *dbg_redzone1(cachep
, objp
),
2065 *dbg_redzone2(cachep
, objp
));
2068 if (cachep
->flags
& SLAB_STORE_USER
) {
2069 printk(KERN_ERR
"Last user: [<%p>]",
2070 *dbg_userword(cachep
, objp
));
2071 print_symbol("(%s)",
2072 (unsigned long)*dbg_userword(cachep
, objp
));
2075 realobj
= (char *)objp
+ obj_offset(cachep
);
2076 size
= cachep
->object_size
;
2077 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
2080 if (i
+ limit
> size
)
2082 dump_line(realobj
, i
, limit
);
2086 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
2092 realobj
= (char *)objp
+ obj_offset(cachep
);
2093 size
= cachep
->object_size
;
2095 for (i
= 0; i
< size
; i
++) {
2096 char exp
= POISON_FREE
;
2099 if (realobj
[i
] != exp
) {
2105 "Slab corruption (%s): %s start=%p, len=%d\n",
2106 print_tainted(), cachep
->name
, realobj
, size
);
2107 print_objinfo(cachep
, objp
, 0);
2109 /* Hexdump the affected line */
2112 if (i
+ limit
> size
)
2114 dump_line(realobj
, i
, limit
);
2117 /* Limit to 5 lines */
2123 /* Print some data about the neighboring objects, if they
2126 struct slab
*slabp
= virt_to_slab(objp
);
2129 objnr
= obj_to_index(cachep
, slabp
, objp
);
2131 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
2132 realobj
= (char *)objp
+ obj_offset(cachep
);
2133 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
2135 print_objinfo(cachep
, objp
, 2);
2137 if (objnr
+ 1 < cachep
->num
) {
2138 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
2139 realobj
= (char *)objp
+ obj_offset(cachep
);
2140 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
2142 print_objinfo(cachep
, objp
, 2);
2149 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2152 for (i
= 0; i
< cachep
->num
; i
++) {
2153 void *objp
= index_to_obj(cachep
, slabp
, i
);
2155 if (cachep
->flags
& SLAB_POISON
) {
2156 #ifdef CONFIG_DEBUG_PAGEALLOC
2157 if (cachep
->size
% PAGE_SIZE
== 0 &&
2159 kernel_map_pages(virt_to_page(objp
),
2160 cachep
->size
/ PAGE_SIZE
, 1);
2162 check_poison_obj(cachep
, objp
);
2164 check_poison_obj(cachep
, objp
);
2167 if (cachep
->flags
& SLAB_RED_ZONE
) {
2168 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2169 slab_error(cachep
, "start of a freed object "
2171 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2172 slab_error(cachep
, "end of a freed object "
2178 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2184 * slab_destroy - destroy and release all objects in a slab
2185 * @cachep: cache pointer being destroyed
2186 * @slabp: slab pointer being destroyed
2188 * Destroy all the objs in a slab, and release the mem back to the system.
2189 * Before calling the slab must have been unlinked from the cache. The
2190 * cache-lock is not held/needed.
2192 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2194 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2196 slab_destroy_debugcheck(cachep
, slabp
);
2197 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2198 struct slab_rcu
*slab_rcu
;
2200 slab_rcu
= (struct slab_rcu
*)slabp
;
2201 slab_rcu
->cachep
= cachep
;
2202 slab_rcu
->addr
= addr
;
2203 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2205 kmem_freepages(cachep
, addr
);
2206 if (OFF_SLAB(cachep
))
2207 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2212 * calculate_slab_order - calculate size (page order) of slabs
2213 * @cachep: pointer to the cache that is being created
2214 * @size: size of objects to be created in this cache.
2215 * @align: required alignment for the objects.
2216 * @flags: slab allocation flags
2218 * Also calculates the number of objects per slab.
2220 * This could be made much more intelligent. For now, try to avoid using
2221 * high order pages for slabs. When the gfp() functions are more friendly
2222 * towards high-order requests, this should be changed.
2224 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2225 size_t size
, size_t align
, unsigned long flags
)
2227 unsigned long offslab_limit
;
2228 size_t left_over
= 0;
2231 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2235 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2239 if (flags
& CFLGS_OFF_SLAB
) {
2241 * Max number of objs-per-slab for caches which
2242 * use off-slab slabs. Needed to avoid a possible
2243 * looping condition in cache_grow().
2245 offslab_limit
= size
- sizeof(struct slab
);
2246 offslab_limit
/= sizeof(kmem_bufctl_t
);
2248 if (num
> offslab_limit
)
2252 /* Found something acceptable - save it away */
2254 cachep
->gfporder
= gfporder
;
2255 left_over
= remainder
;
2258 * A VFS-reclaimable slab tends to have most allocations
2259 * as GFP_NOFS and we really don't want to have to be allocating
2260 * higher-order pages when we are unable to shrink dcache.
2262 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2266 * Large number of objects is good, but very large slabs are
2267 * currently bad for the gfp()s.
2269 if (gfporder
>= slab_max_order
)
2273 * Acceptable internal fragmentation?
2275 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2281 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2283 if (slab_state
>= FULL
)
2284 return enable_cpucache(cachep
, gfp
);
2286 if (slab_state
== DOWN
) {
2288 * Note: the first kmem_cache_create must create the cache
2289 * that's used by kmalloc(24), otherwise the creation of
2290 * further caches will BUG().
2292 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2295 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2296 * the first cache, then we need to set up all its list3s,
2297 * otherwise the creation of further caches will BUG().
2299 set_up_list3s(cachep
, SIZE_AC
);
2300 if (INDEX_AC
== INDEX_L3
)
2301 slab_state
= PARTIAL_L3
;
2303 slab_state
= PARTIAL_ARRAYCACHE
;
2305 cachep
->array
[smp_processor_id()] =
2306 kmalloc(sizeof(struct arraycache_init
), gfp
);
2308 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2309 set_up_list3s(cachep
, SIZE_L3
);
2310 slab_state
= PARTIAL_L3
;
2313 for_each_online_node(node
) {
2314 cachep
->nodelists
[node
] =
2315 kmalloc_node(sizeof(struct kmem_list3
),
2317 BUG_ON(!cachep
->nodelists
[node
]);
2318 kmem_list3_init(cachep
->nodelists
[node
]);
2322 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2323 jiffies
+ REAPTIMEOUT_LIST3
+
2324 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2326 cpu_cache_get(cachep
)->avail
= 0;
2327 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2328 cpu_cache_get(cachep
)->batchcount
= 1;
2329 cpu_cache_get(cachep
)->touched
= 0;
2330 cachep
->batchcount
= 1;
2331 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2336 * __kmem_cache_create - Create a cache.
2337 * @name: A string which is used in /proc/slabinfo to identify this cache.
2338 * @size: The size of objects to be created in this cache.
2339 * @align: The required alignment for the objects.
2340 * @flags: SLAB flags
2341 * @ctor: A constructor for the objects.
2343 * Returns a ptr to the cache on success, NULL on failure.
2344 * Cannot be called within a int, but can be interrupted.
2345 * The @ctor is run when new pages are allocated by the cache.
2349 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2350 * to catch references to uninitialised memory.
2352 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2353 * for buffer overruns.
2355 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2356 * cacheline. This can be beneficial if you're counting cycles as closely
2360 __kmem_cache_create (const char *name
, size_t size
, size_t align
,
2361 unsigned long flags
, void (*ctor
)(void *))
2363 size_t left_over
, slab_size
, ralign
;
2364 struct kmem_cache
*cachep
= NULL
;
2370 * Enable redzoning and last user accounting, except for caches with
2371 * large objects, if the increased size would increase the object size
2372 * above the next power of two: caches with object sizes just above a
2373 * power of two have a significant amount of internal fragmentation.
2375 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2376 2 * sizeof(unsigned long long)))
2377 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2378 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2379 flags
|= SLAB_POISON
;
2381 if (flags
& SLAB_DESTROY_BY_RCU
)
2382 BUG_ON(flags
& SLAB_POISON
);
2385 * Always checks flags, a caller might be expecting debug support which
2388 BUG_ON(flags
& ~CREATE_MASK
);
2391 * Check that size is in terms of words. This is needed to avoid
2392 * unaligned accesses for some archs when redzoning is used, and makes
2393 * sure any on-slab bufctl's are also correctly aligned.
2395 if (size
& (BYTES_PER_WORD
- 1)) {
2396 size
+= (BYTES_PER_WORD
- 1);
2397 size
&= ~(BYTES_PER_WORD
- 1);
2400 /* calculate the final buffer alignment: */
2402 /* 1) arch recommendation: can be overridden for debug */
2403 if (flags
& SLAB_HWCACHE_ALIGN
) {
2405 * Default alignment: as specified by the arch code. Except if
2406 * an object is really small, then squeeze multiple objects into
2409 ralign
= cache_line_size();
2410 while (size
<= ralign
/ 2)
2413 ralign
= BYTES_PER_WORD
;
2417 * Redzoning and user store require word alignment or possibly larger.
2418 * Note this will be overridden by architecture or caller mandated
2419 * alignment if either is greater than BYTES_PER_WORD.
2421 if (flags
& SLAB_STORE_USER
)
2422 ralign
= BYTES_PER_WORD
;
2424 if (flags
& SLAB_RED_ZONE
) {
2425 ralign
= REDZONE_ALIGN
;
2426 /* If redzoning, ensure that the second redzone is suitably
2427 * aligned, by adjusting the object size accordingly. */
2428 size
+= REDZONE_ALIGN
- 1;
2429 size
&= ~(REDZONE_ALIGN
- 1);
2432 /* 2) arch mandated alignment */
2433 if (ralign
< ARCH_SLAB_MINALIGN
) {
2434 ralign
= ARCH_SLAB_MINALIGN
;
2436 /* 3) caller mandated alignment */
2437 if (ralign
< align
) {
2440 /* disable debug if necessary */
2441 if (ralign
> __alignof__(unsigned long long))
2442 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2448 if (slab_is_available())
2453 /* Get cache's description obj. */
2454 cachep
= kmem_cache_zalloc(kmem_cache
, gfp
);
2458 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
2459 cachep
->object_size
= size
;
2460 cachep
->align
= align
;
2464 * Both debugging options require word-alignment which is calculated
2467 if (flags
& SLAB_RED_ZONE
) {
2468 /* add space for red zone words */
2469 cachep
->obj_offset
+= sizeof(unsigned long long);
2470 size
+= 2 * sizeof(unsigned long long);
2472 if (flags
& SLAB_STORE_USER
) {
2473 /* user store requires one word storage behind the end of
2474 * the real object. But if the second red zone needs to be
2475 * aligned to 64 bits, we must allow that much space.
2477 if (flags
& SLAB_RED_ZONE
)
2478 size
+= REDZONE_ALIGN
;
2480 size
+= BYTES_PER_WORD
;
2482 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2483 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2484 && cachep
->object_size
> cache_line_size() && ALIGN(size
, align
) < PAGE_SIZE
) {
2485 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, align
);
2492 * Determine if the slab management is 'on' or 'off' slab.
2493 * (bootstrapping cannot cope with offslab caches so don't do
2494 * it too early on. Always use on-slab management when
2495 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2497 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2498 !(flags
& SLAB_NOLEAKTRACE
))
2500 * Size is large, assume best to place the slab management obj
2501 * off-slab (should allow better packing of objs).
2503 flags
|= CFLGS_OFF_SLAB
;
2505 size
= ALIGN(size
, align
);
2507 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2511 "kmem_cache_create: couldn't create cache %s.\n", name
);
2512 kmem_cache_free(kmem_cache
, cachep
);
2515 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2516 + sizeof(struct slab
), align
);
2519 * If the slab has been placed off-slab, and we have enough space then
2520 * move it on-slab. This is at the expense of any extra colouring.
2522 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2523 flags
&= ~CFLGS_OFF_SLAB
;
2524 left_over
-= slab_size
;
2527 if (flags
& CFLGS_OFF_SLAB
) {
2528 /* really off slab. No need for manual alignment */
2530 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2532 #ifdef CONFIG_PAGE_POISONING
2533 /* If we're going to use the generic kernel_map_pages()
2534 * poisoning, then it's going to smash the contents of
2535 * the redzone and userword anyhow, so switch them off.
2537 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2538 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2542 cachep
->colour_off
= cache_line_size();
2543 /* Offset must be a multiple of the alignment. */
2544 if (cachep
->colour_off
< align
)
2545 cachep
->colour_off
= align
;
2546 cachep
->colour
= left_over
/ cachep
->colour_off
;
2547 cachep
->slab_size
= slab_size
;
2548 cachep
->flags
= flags
;
2549 cachep
->allocflags
= 0;
2550 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2551 cachep
->allocflags
|= GFP_DMA
;
2552 cachep
->size
= size
;
2553 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2555 if (flags
& CFLGS_OFF_SLAB
) {
2556 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2558 * This is a possibility for one of the malloc_sizes caches.
2559 * But since we go off slab only for object size greater than
2560 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2561 * this should not happen at all.
2562 * But leave a BUG_ON for some lucky dude.
2564 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2566 cachep
->ctor
= ctor
;
2567 cachep
->name
= name
;
2568 cachep
->refcount
= 1;
2570 if (setup_cpu_cache(cachep
, gfp
)) {
2571 __kmem_cache_shutdown(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
);
2589 static void check_irq_off(void)
2591 BUG_ON(!irqs_disabled());
2594 static void check_irq_on(void)
2596 BUG_ON(irqs_disabled());
2599 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2603 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2607 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2611 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2616 #define check_irq_off() do { } while(0)
2617 #define check_irq_on() do { } while(0)
2618 #define check_spinlock_acquired(x) do { } while(0)
2619 #define check_spinlock_acquired_node(x, y) do { } while(0)
2622 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2623 struct array_cache
*ac
,
2624 int force
, int node
);
2626 static void do_drain(void *arg
)
2628 struct kmem_cache
*cachep
= arg
;
2629 struct array_cache
*ac
;
2630 int node
= numa_mem_id();
2633 ac
= cpu_cache_get(cachep
);
2634 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2635 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2636 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2640 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2642 struct kmem_list3
*l3
;
2645 on_each_cpu(do_drain
, cachep
, 1);
2647 for_each_online_node(node
) {
2648 l3
= cachep
->nodelists
[node
];
2649 if (l3
&& l3
->alien
)
2650 drain_alien_cache(cachep
, l3
->alien
);
2653 for_each_online_node(node
) {
2654 l3
= cachep
->nodelists
[node
];
2656 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2661 * Remove slabs from the list of free slabs.
2662 * Specify the number of slabs to drain in tofree.
2664 * Returns the actual number of slabs released.
2666 static int drain_freelist(struct kmem_cache
*cache
,
2667 struct kmem_list3
*l3
, int tofree
)
2669 struct list_head
*p
;
2674 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2676 spin_lock_irq(&l3
->list_lock
);
2677 p
= l3
->slabs_free
.prev
;
2678 if (p
== &l3
->slabs_free
) {
2679 spin_unlock_irq(&l3
->list_lock
);
2683 slabp
= list_entry(p
, struct slab
, list
);
2685 BUG_ON(slabp
->inuse
);
2687 list_del(&slabp
->list
);
2689 * Safe to drop the lock. The slab is no longer linked
2692 l3
->free_objects
-= cache
->num
;
2693 spin_unlock_irq(&l3
->list_lock
);
2694 slab_destroy(cache
, slabp
);
2701 /* Called with slab_mutex held to protect against cpu hotplug */
2702 static int __cache_shrink(struct kmem_cache
*cachep
)
2705 struct kmem_list3
*l3
;
2707 drain_cpu_caches(cachep
);
2710 for_each_online_node(i
) {
2711 l3
= cachep
->nodelists
[i
];
2715 drain_freelist(cachep
, l3
, l3
->free_objects
);
2717 ret
+= !list_empty(&l3
->slabs_full
) ||
2718 !list_empty(&l3
->slabs_partial
);
2720 return (ret
? 1 : 0);
2724 * kmem_cache_shrink - Shrink a cache.
2725 * @cachep: The cache to shrink.
2727 * Releases as many slabs as possible for a cache.
2728 * To help debugging, a zero exit status indicates all slabs were released.
2730 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2733 BUG_ON(!cachep
|| in_interrupt());
2736 mutex_lock(&slab_mutex
);
2737 ret
= __cache_shrink(cachep
);
2738 mutex_unlock(&slab_mutex
);
2742 EXPORT_SYMBOL(kmem_cache_shrink
);
2744 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2747 struct kmem_list3
*l3
;
2748 int rc
= __cache_shrink(cachep
);
2753 for_each_online_cpu(i
)
2754 kfree(cachep
->array
[i
]);
2756 /* NUMA: free the list3 structures */
2757 for_each_online_node(i
) {
2758 l3
= cachep
->nodelists
[i
];
2761 free_alien_cache(l3
->alien
);
2769 * Get the memory for a slab management obj.
2770 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2771 * always come from malloc_sizes caches. The slab descriptor cannot
2772 * come from the same cache which is getting created because,
2773 * when we are searching for an appropriate cache for these
2774 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2775 * If we are creating a malloc_sizes cache here it would not be visible to
2776 * kmem_find_general_cachep till the initialization is complete.
2777 * Hence we cannot have slabp_cache same as the original cache.
2779 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2780 int colour_off
, gfp_t local_flags
,
2785 if (OFF_SLAB(cachep
)) {
2786 /* Slab management obj is off-slab. */
2787 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2788 local_flags
, nodeid
);
2790 * If the first object in the slab is leaked (it's allocated
2791 * but no one has a reference to it), we want to make sure
2792 * kmemleak does not treat the ->s_mem pointer as a reference
2793 * to the object. Otherwise we will not report the leak.
2795 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2800 slabp
= objp
+ colour_off
;
2801 colour_off
+= cachep
->slab_size
;
2804 slabp
->colouroff
= colour_off
;
2805 slabp
->s_mem
= objp
+ colour_off
;
2806 slabp
->nodeid
= nodeid
;
2811 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2813 return (kmem_bufctl_t
*) (slabp
+ 1);
2816 static void cache_init_objs(struct kmem_cache
*cachep
,
2821 for (i
= 0; i
< cachep
->num
; i
++) {
2822 void *objp
= index_to_obj(cachep
, slabp
, i
);
2824 /* need to poison the objs? */
2825 if (cachep
->flags
& SLAB_POISON
)
2826 poison_obj(cachep
, objp
, POISON_FREE
);
2827 if (cachep
->flags
& SLAB_STORE_USER
)
2828 *dbg_userword(cachep
, objp
) = NULL
;
2830 if (cachep
->flags
& SLAB_RED_ZONE
) {
2831 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2832 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2835 * Constructors are not allowed to allocate memory from the same
2836 * cache which they are a constructor for. Otherwise, deadlock.
2837 * They must also be threaded.
2839 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2840 cachep
->ctor(objp
+ obj_offset(cachep
));
2842 if (cachep
->flags
& SLAB_RED_ZONE
) {
2843 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2844 slab_error(cachep
, "constructor overwrote the"
2845 " end of an object");
2846 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2847 slab_error(cachep
, "constructor overwrote the"
2848 " start of an object");
2850 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2851 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2852 kernel_map_pages(virt_to_page(objp
),
2853 cachep
->size
/ PAGE_SIZE
, 0);
2858 slab_bufctl(slabp
)[i
] = i
+ 1;
2860 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2863 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2865 if (CONFIG_ZONE_DMA_FLAG
) {
2866 if (flags
& GFP_DMA
)
2867 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2869 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2873 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2876 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2880 next
= slab_bufctl(slabp
)[slabp
->free
];
2882 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2883 WARN_ON(slabp
->nodeid
!= nodeid
);
2890 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2891 void *objp
, int nodeid
)
2893 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2896 /* Verify that the slab belongs to the intended node */
2897 WARN_ON(slabp
->nodeid
!= nodeid
);
2899 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2900 printk(KERN_ERR
"slab: double free detected in cache "
2901 "'%s', objp %p\n", cachep
->name
, objp
);
2905 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2906 slabp
->free
= objnr
;
2911 * Map pages beginning at addr to the given cache and slab. This is required
2912 * for the slab allocator to be able to lookup the cache and slab of a
2913 * virtual address for kfree, ksize, and slab debugging.
2915 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2921 page
= virt_to_page(addr
);
2924 if (likely(!PageCompound(page
)))
2925 nr_pages
<<= cache
->gfporder
;
2928 page
->slab_cache
= cache
;
2929 page
->slab_page
= slab
;
2931 } while (--nr_pages
);
2935 * Grow (by 1) the number of slabs within a cache. This is called by
2936 * kmem_cache_alloc() when there are no active objs left in a cache.
2938 static int cache_grow(struct kmem_cache
*cachep
,
2939 gfp_t flags
, int nodeid
, void *objp
)
2944 struct kmem_list3
*l3
;
2947 * Be lazy and only check for valid flags here, keeping it out of the
2948 * critical path in kmem_cache_alloc().
2950 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2951 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2953 /* Take the l3 list lock to change the colour_next on this node */
2955 l3
= cachep
->nodelists
[nodeid
];
2956 spin_lock(&l3
->list_lock
);
2958 /* Get colour for the slab, and cal the next value. */
2959 offset
= l3
->colour_next
;
2961 if (l3
->colour_next
>= cachep
->colour
)
2962 l3
->colour_next
= 0;
2963 spin_unlock(&l3
->list_lock
);
2965 offset
*= cachep
->colour_off
;
2967 if (local_flags
& __GFP_WAIT
)
2971 * The test for missing atomic flag is performed here, rather than
2972 * the more obvious place, simply to reduce the critical path length
2973 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2974 * will eventually be caught here (where it matters).
2976 kmem_flagcheck(cachep
, flags
);
2979 * Get mem for the objs. Attempt to allocate a physical page from
2983 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2987 /* Get slab management. */
2988 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2989 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2993 slab_map_pages(cachep
, slabp
, objp
);
2995 cache_init_objs(cachep
, slabp
);
2997 if (local_flags
& __GFP_WAIT
)
2998 local_irq_disable();
3000 spin_lock(&l3
->list_lock
);
3002 /* Make slab active. */
3003 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
3004 STATS_INC_GROWN(cachep
);
3005 l3
->free_objects
+= cachep
->num
;
3006 spin_unlock(&l3
->list_lock
);
3009 kmem_freepages(cachep
, objp
);
3011 if (local_flags
& __GFP_WAIT
)
3012 local_irq_disable();
3019 * Perform extra freeing checks:
3020 * - detect bad pointers.
3021 * - POISON/RED_ZONE checking
3023 static void kfree_debugcheck(const void *objp
)
3025 if (!virt_addr_valid(objp
)) {
3026 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
3027 (unsigned long)objp
);
3032 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
3034 unsigned long long redzone1
, redzone2
;
3036 redzone1
= *dbg_redzone1(cache
, obj
);
3037 redzone2
= *dbg_redzone2(cache
, obj
);
3042 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
3045 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
3046 slab_error(cache
, "double free detected");
3048 slab_error(cache
, "memory outside object was overwritten");
3050 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3051 obj
, redzone1
, redzone2
);
3054 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
3061 BUG_ON(virt_to_cache(objp
) != cachep
);
3063 objp
-= obj_offset(cachep
);
3064 kfree_debugcheck(objp
);
3065 page
= virt_to_head_page(objp
);
3067 slabp
= page
->slab_page
;
3069 if (cachep
->flags
& SLAB_RED_ZONE
) {
3070 verify_redzone_free(cachep
, objp
);
3071 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
3072 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
3074 if (cachep
->flags
& SLAB_STORE_USER
)
3075 *dbg_userword(cachep
, objp
) = caller
;
3077 objnr
= obj_to_index(cachep
, slabp
, objp
);
3079 BUG_ON(objnr
>= cachep
->num
);
3080 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
3082 #ifdef CONFIG_DEBUG_SLAB_LEAK
3083 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
3085 if (cachep
->flags
& SLAB_POISON
) {
3086 #ifdef CONFIG_DEBUG_PAGEALLOC
3087 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
3088 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
3089 kernel_map_pages(virt_to_page(objp
),
3090 cachep
->size
/ PAGE_SIZE
, 0);
3092 poison_obj(cachep
, objp
, POISON_FREE
);
3095 poison_obj(cachep
, objp
, POISON_FREE
);
3101 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
3106 /* Check slab's freelist to see if this obj is there. */
3107 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
3109 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3112 if (entries
!= cachep
->num
- slabp
->inuse
) {
3114 printk(KERN_ERR
"slab: Internal list corruption detected in "
3115 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3116 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
3118 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3119 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3125 #define kfree_debugcheck(x) do { } while(0)
3126 #define cache_free_debugcheck(x,objp,z) (objp)
3127 #define check_slabp(x,y) do { } while(0)
3130 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
3134 struct kmem_list3
*l3
;
3135 struct array_cache
*ac
;
3139 node
= numa_mem_id();
3140 if (unlikely(force_refill
))
3143 ac
= cpu_cache_get(cachep
);
3144 batchcount
= ac
->batchcount
;
3145 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3147 * If there was little recent activity on this cache, then
3148 * perform only a partial refill. Otherwise we could generate
3151 batchcount
= BATCHREFILL_LIMIT
;
3153 l3
= cachep
->nodelists
[node
];
3155 BUG_ON(ac
->avail
> 0 || !l3
);
3156 spin_lock(&l3
->list_lock
);
3158 /* See if we can refill from the shared array */
3159 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3160 l3
->shared
->touched
= 1;
3164 while (batchcount
> 0) {
3165 struct list_head
*entry
;
3167 /* Get slab alloc is to come from. */
3168 entry
= l3
->slabs_partial
.next
;
3169 if (entry
== &l3
->slabs_partial
) {
3170 l3
->free_touched
= 1;
3171 entry
= l3
->slabs_free
.next
;
3172 if (entry
== &l3
->slabs_free
)
3176 slabp
= list_entry(entry
, struct slab
, list
);
3177 check_slabp(cachep
, slabp
);
3178 check_spinlock_acquired(cachep
);
3181 * The slab was either on partial or free list so
3182 * there must be at least one object available for
3185 BUG_ON(slabp
->inuse
>= cachep
->num
);
3187 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3188 STATS_INC_ALLOCED(cachep
);
3189 STATS_INC_ACTIVE(cachep
);
3190 STATS_SET_HIGH(cachep
);
3192 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
3195 check_slabp(cachep
, slabp
);
3197 /* move slabp to correct slabp list: */
3198 list_del(&slabp
->list
);
3199 if (slabp
->free
== BUFCTL_END
)
3200 list_add(&slabp
->list
, &l3
->slabs_full
);
3202 list_add(&slabp
->list
, &l3
->slabs_partial
);
3206 l3
->free_objects
-= ac
->avail
;
3208 spin_unlock(&l3
->list_lock
);
3210 if (unlikely(!ac
->avail
)) {
3213 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3215 /* cache_grow can reenable interrupts, then ac could change. */
3216 ac
= cpu_cache_get(cachep
);
3218 /* no objects in sight? abort */
3219 if (!x
&& (ac
->avail
== 0 || force_refill
))
3222 if (!ac
->avail
) /* objects refilled by interrupt? */
3227 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
3230 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3233 might_sleep_if(flags
& __GFP_WAIT
);
3235 kmem_flagcheck(cachep
, flags
);
3240 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3241 gfp_t flags
, void *objp
, void *caller
)
3245 if (cachep
->flags
& SLAB_POISON
) {
3246 #ifdef CONFIG_DEBUG_PAGEALLOC
3247 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3248 kernel_map_pages(virt_to_page(objp
),
3249 cachep
->size
/ PAGE_SIZE
, 1);
3251 check_poison_obj(cachep
, objp
);
3253 check_poison_obj(cachep
, objp
);
3255 poison_obj(cachep
, objp
, POISON_INUSE
);
3257 if (cachep
->flags
& SLAB_STORE_USER
)
3258 *dbg_userword(cachep
, objp
) = caller
;
3260 if (cachep
->flags
& SLAB_RED_ZONE
) {
3261 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3262 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3263 slab_error(cachep
, "double free, or memory outside"
3264 " object was overwritten");
3266 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3267 objp
, *dbg_redzone1(cachep
, objp
),
3268 *dbg_redzone2(cachep
, objp
));
3270 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3271 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3273 #ifdef CONFIG_DEBUG_SLAB_LEAK
3278 slabp
= virt_to_head_page(objp
)->slab_page
;
3279 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3280 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3283 objp
+= obj_offset(cachep
);
3284 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3286 if (ARCH_SLAB_MINALIGN
&&
3287 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3288 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3289 objp
, (int)ARCH_SLAB_MINALIGN
);
3294 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3297 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3299 if (cachep
== kmem_cache
)
3302 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3305 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3308 struct array_cache
*ac
;
3309 bool force_refill
= false;
3313 ac
= cpu_cache_get(cachep
);
3314 if (likely(ac
->avail
)) {
3316 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3319 * Allow for the possibility all avail objects are not allowed
3320 * by the current flags
3323 STATS_INC_ALLOCHIT(cachep
);
3326 force_refill
= true;
3329 STATS_INC_ALLOCMISS(cachep
);
3330 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3332 * the 'ac' may be updated by cache_alloc_refill(),
3333 * and kmemleak_erase() requires its correct value.
3335 ac
= cpu_cache_get(cachep
);
3339 * To avoid a false negative, if an object that is in one of the
3340 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3341 * treat the array pointers as a reference to the object.
3344 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3350 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3352 * If we are in_interrupt, then process context, including cpusets and
3353 * mempolicy, may not apply and should not be used for allocation policy.
3355 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3357 int nid_alloc
, nid_here
;
3359 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3361 nid_alloc
= nid_here
= numa_mem_id();
3362 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3363 nid_alloc
= cpuset_slab_spread_node();
3364 else if (current
->mempolicy
)
3365 nid_alloc
= slab_node();
3366 if (nid_alloc
!= nid_here
)
3367 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3372 * Fallback function if there was no memory available and no objects on a
3373 * certain node and fall back is permitted. First we scan all the
3374 * available nodelists for available objects. If that fails then we
3375 * perform an allocation without specifying a node. This allows the page
3376 * allocator to do its reclaim / fallback magic. We then insert the
3377 * slab into the proper nodelist and then allocate from it.
3379 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3381 struct zonelist
*zonelist
;
3385 enum zone_type high_zoneidx
= gfp_zone(flags
);
3388 unsigned int cpuset_mems_cookie
;
3390 if (flags
& __GFP_THISNODE
)
3393 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3396 cpuset_mems_cookie
= get_mems_allowed();
3397 zonelist
= node_zonelist(slab_node(), flags
);
3401 * Look through allowed nodes for objects available
3402 * from existing per node queues.
3404 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3405 nid
= zone_to_nid(zone
);
3407 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3408 cache
->nodelists
[nid
] &&
3409 cache
->nodelists
[nid
]->free_objects
) {
3410 obj
= ____cache_alloc_node(cache
,
3411 flags
| GFP_THISNODE
, nid
);
3419 * This allocation will be performed within the constraints
3420 * of the current cpuset / memory policy requirements.
3421 * We may trigger various forms of reclaim on the allowed
3422 * set and go into memory reserves if necessary.
3424 if (local_flags
& __GFP_WAIT
)
3426 kmem_flagcheck(cache
, flags
);
3427 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3428 if (local_flags
& __GFP_WAIT
)
3429 local_irq_disable();
3432 * Insert into the appropriate per node queues
3434 nid
= page_to_nid(virt_to_page(obj
));
3435 if (cache_grow(cache
, flags
, nid
, obj
)) {
3436 obj
= ____cache_alloc_node(cache
,
3437 flags
| GFP_THISNODE
, nid
);
3440 * Another processor may allocate the
3441 * objects in the slab since we are
3442 * not holding any locks.
3446 /* cache_grow already freed obj */
3452 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3458 * A interface to enable slab creation on nodeid
3460 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3463 struct list_head
*entry
;
3465 struct kmem_list3
*l3
;
3469 l3
= cachep
->nodelists
[nodeid
];
3474 spin_lock(&l3
->list_lock
);
3475 entry
= l3
->slabs_partial
.next
;
3476 if (entry
== &l3
->slabs_partial
) {
3477 l3
->free_touched
= 1;
3478 entry
= l3
->slabs_free
.next
;
3479 if (entry
== &l3
->slabs_free
)
3483 slabp
= list_entry(entry
, struct slab
, list
);
3484 check_spinlock_acquired_node(cachep
, nodeid
);
3485 check_slabp(cachep
, slabp
);
3487 STATS_INC_NODEALLOCS(cachep
);
3488 STATS_INC_ACTIVE(cachep
);
3489 STATS_SET_HIGH(cachep
);
3491 BUG_ON(slabp
->inuse
== cachep
->num
);
3493 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3494 check_slabp(cachep
, slabp
);
3496 /* move slabp to correct slabp list: */
3497 list_del(&slabp
->list
);
3499 if (slabp
->free
== BUFCTL_END
)
3500 list_add(&slabp
->list
, &l3
->slabs_full
);
3502 list_add(&slabp
->list
, &l3
->slabs_partial
);
3504 spin_unlock(&l3
->list_lock
);
3508 spin_unlock(&l3
->list_lock
);
3509 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3513 return fallback_alloc(cachep
, flags
);
3520 * kmem_cache_alloc_node - Allocate an object on the specified node
3521 * @cachep: The cache to allocate from.
3522 * @flags: See kmalloc().
3523 * @nodeid: node number of the target node.
3524 * @caller: return address of caller, used for debug information
3526 * Identical to kmem_cache_alloc but it will allocate memory on the given
3527 * node, which can improve the performance for cpu bound structures.
3529 * Fallback to other node is possible if __GFP_THISNODE is not set.
3531 static __always_inline
void *
3532 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3535 unsigned long save_flags
;
3537 int slab_node
= numa_mem_id();
3539 flags
&= gfp_allowed_mask
;
3541 lockdep_trace_alloc(flags
);
3543 if (slab_should_failslab(cachep
, flags
))
3546 cache_alloc_debugcheck_before(cachep
, flags
);
3547 local_irq_save(save_flags
);
3549 if (nodeid
== NUMA_NO_NODE
)
3552 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3553 /* Node not bootstrapped yet */
3554 ptr
= fallback_alloc(cachep
, flags
);
3558 if (nodeid
== slab_node
) {
3560 * Use the locally cached objects if possible.
3561 * However ____cache_alloc does not allow fallback
3562 * to other nodes. It may fail while we still have
3563 * objects on other nodes available.
3565 ptr
= ____cache_alloc(cachep
, flags
);
3569 /* ___cache_alloc_node can fall back to other nodes */
3570 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3572 local_irq_restore(save_flags
);
3573 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3574 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3578 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3580 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3581 memset(ptr
, 0, cachep
->object_size
);
3586 static __always_inline
void *
3587 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3591 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3592 objp
= alternate_node_alloc(cache
, flags
);
3596 objp
= ____cache_alloc(cache
, flags
);
3599 * We may just have run out of memory on the local node.
3600 * ____cache_alloc_node() knows how to locate memory on other nodes
3603 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3610 static __always_inline
void *
3611 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3613 return ____cache_alloc(cachep
, flags
);
3616 #endif /* CONFIG_NUMA */
3618 static __always_inline
void *
3619 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3621 unsigned long save_flags
;
3624 flags
&= gfp_allowed_mask
;
3626 lockdep_trace_alloc(flags
);
3628 if (slab_should_failslab(cachep
, flags
))
3631 cache_alloc_debugcheck_before(cachep
, flags
);
3632 local_irq_save(save_flags
);
3633 objp
= __do_cache_alloc(cachep
, flags
);
3634 local_irq_restore(save_flags
);
3635 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3636 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3641 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3643 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3644 memset(objp
, 0, cachep
->object_size
);
3650 * Caller needs to acquire correct kmem_list's list_lock
3652 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3656 struct kmem_list3
*l3
;
3658 for (i
= 0; i
< nr_objects
; i
++) {
3662 clear_obj_pfmemalloc(&objpp
[i
]);
3665 slabp
= virt_to_slab(objp
);
3666 l3
= cachep
->nodelists
[node
];
3667 list_del(&slabp
->list
);
3668 check_spinlock_acquired_node(cachep
, node
);
3669 check_slabp(cachep
, slabp
);
3670 slab_put_obj(cachep
, slabp
, objp
, node
);
3671 STATS_DEC_ACTIVE(cachep
);
3673 check_slabp(cachep
, slabp
);
3675 /* fixup slab chains */
3676 if (slabp
->inuse
== 0) {
3677 if (l3
->free_objects
> l3
->free_limit
) {
3678 l3
->free_objects
-= cachep
->num
;
3679 /* No need to drop any previously held
3680 * lock here, even if we have a off-slab slab
3681 * descriptor it is guaranteed to come from
3682 * a different cache, refer to comments before
3685 slab_destroy(cachep
, slabp
);
3687 list_add(&slabp
->list
, &l3
->slabs_free
);
3690 /* Unconditionally move a slab to the end of the
3691 * partial list on free - maximum time for the
3692 * other objects to be freed, too.
3694 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3699 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3702 struct kmem_list3
*l3
;
3703 int node
= numa_mem_id();
3705 batchcount
= ac
->batchcount
;
3707 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3710 l3
= cachep
->nodelists
[node
];
3711 spin_lock(&l3
->list_lock
);
3713 struct array_cache
*shared_array
= l3
->shared
;
3714 int max
= shared_array
->limit
- shared_array
->avail
;
3716 if (batchcount
> max
)
3718 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3719 ac
->entry
, sizeof(void *) * batchcount
);
3720 shared_array
->avail
+= batchcount
;
3725 free_block(cachep
, ac
->entry
, batchcount
, node
);
3730 struct list_head
*p
;
3732 p
= l3
->slabs_free
.next
;
3733 while (p
!= &(l3
->slabs_free
)) {
3736 slabp
= list_entry(p
, struct slab
, list
);
3737 BUG_ON(slabp
->inuse
);
3742 STATS_SET_FREEABLE(cachep
, i
);
3745 spin_unlock(&l3
->list_lock
);
3746 ac
->avail
-= batchcount
;
3747 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3751 * Release an obj back to its cache. If the obj has a constructed state, it must
3752 * be in this state _before_ it is released. Called with disabled ints.
3754 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3757 struct array_cache
*ac
= cpu_cache_get(cachep
);
3760 kmemleak_free_recursive(objp
, cachep
->flags
);
3761 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3763 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3766 * Skip calling cache_free_alien() when the platform is not numa.
3767 * This will avoid cache misses that happen while accessing slabp (which
3768 * is per page memory reference) to get nodeid. Instead use a global
3769 * variable to skip the call, which is mostly likely to be present in
3772 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3775 if (likely(ac
->avail
< ac
->limit
)) {
3776 STATS_INC_FREEHIT(cachep
);
3778 STATS_INC_FREEMISS(cachep
);
3779 cache_flusharray(cachep
, ac
);
3782 ac_put_obj(cachep
, ac
, objp
);
3786 * kmem_cache_alloc - Allocate an object
3787 * @cachep: The cache to allocate from.
3788 * @flags: See kmalloc().
3790 * Allocate an object from this cache. The flags are only relevant
3791 * if the cache has no available objects.
3793 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3795 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3797 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3798 cachep
->object_size
, cachep
->size
, flags
);
3802 EXPORT_SYMBOL(kmem_cache_alloc
);
3804 #ifdef CONFIG_TRACING
3806 kmem_cache_alloc_trace(size_t size
, struct kmem_cache
*cachep
, gfp_t flags
)
3810 ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3812 trace_kmalloc(_RET_IP_
, ret
,
3813 size
, slab_buffer_size(cachep
), flags
);
3816 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3820 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3822 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3823 __builtin_return_address(0));
3825 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3826 cachep
->object_size
, cachep
->size
,
3831 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3833 #ifdef CONFIG_TRACING
3834 void *kmem_cache_alloc_node_trace(size_t size
,
3835 struct kmem_cache
*cachep
,
3841 ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3842 __builtin_return_address(0));
3843 trace_kmalloc_node(_RET_IP_
, ret
,
3844 size
, slab_buffer_size(cachep
),
3848 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3851 static __always_inline
void *
3852 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3854 struct kmem_cache
*cachep
;
3856 cachep
= kmem_find_general_cachep(size
, flags
);
3857 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3859 return kmem_cache_alloc_node_trace(size
, cachep
, flags
, node
);
3862 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3863 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3865 return __do_kmalloc_node(size
, flags
, node
,
3866 __builtin_return_address(0));
3868 EXPORT_SYMBOL(__kmalloc_node
);
3870 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3871 int node
, unsigned long caller
)
3873 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3875 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3877 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3879 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3881 EXPORT_SYMBOL(__kmalloc_node
);
3882 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3883 #endif /* CONFIG_NUMA */
3886 * __do_kmalloc - allocate memory
3887 * @size: how many bytes of memory are required.
3888 * @flags: the type of memory to allocate (see kmalloc).
3889 * @caller: function caller for debug tracking of the caller
3891 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3894 struct kmem_cache
*cachep
;
3897 /* If you want to save a few bytes .text space: replace
3899 * Then kmalloc uses the uninlined functions instead of the inline
3902 cachep
= __find_general_cachep(size
, flags
);
3903 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3905 ret
= __cache_alloc(cachep
, flags
, caller
);
3907 trace_kmalloc((unsigned long) caller
, ret
,
3908 size
, cachep
->size
, flags
);
3914 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3915 void *__kmalloc(size_t size
, gfp_t flags
)
3917 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3919 EXPORT_SYMBOL(__kmalloc
);
3921 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3923 return __do_kmalloc(size
, flags
, (void *)caller
);
3925 EXPORT_SYMBOL(__kmalloc_track_caller
);
3928 void *__kmalloc(size_t size
, gfp_t flags
)
3930 return __do_kmalloc(size
, flags
, NULL
);
3932 EXPORT_SYMBOL(__kmalloc
);
3936 * kmem_cache_free - Deallocate an object
3937 * @cachep: The cache the allocation was from.
3938 * @objp: The previously allocated object.
3940 * Free an object which was previously allocated from this
3943 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3945 unsigned long flags
;
3947 local_irq_save(flags
);
3948 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3949 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3950 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3951 __cache_free(cachep
, objp
, __builtin_return_address(0));
3952 local_irq_restore(flags
);
3954 trace_kmem_cache_free(_RET_IP_
, objp
);
3956 EXPORT_SYMBOL(kmem_cache_free
);
3959 * kfree - free previously allocated memory
3960 * @objp: pointer returned by kmalloc.
3962 * If @objp is NULL, no operation is performed.
3964 * Don't free memory not originally allocated by kmalloc()
3965 * or you will run into trouble.
3967 void kfree(const void *objp
)
3969 struct kmem_cache
*c
;
3970 unsigned long flags
;
3972 trace_kfree(_RET_IP_
, objp
);
3974 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3976 local_irq_save(flags
);
3977 kfree_debugcheck(objp
);
3978 c
= virt_to_cache(objp
);
3979 debug_check_no_locks_freed(objp
, c
->object_size
);
3981 debug_check_no_obj_freed(objp
, c
->object_size
);
3982 __cache_free(c
, (void *)objp
, __builtin_return_address(0));
3983 local_irq_restore(flags
);
3985 EXPORT_SYMBOL(kfree
);
3987 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3989 return cachep
->object_size
;
3991 EXPORT_SYMBOL(kmem_cache_size
);
3994 * This initializes kmem_list3 or resizes various caches for all nodes.
3996 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3999 struct kmem_list3
*l3
;
4000 struct array_cache
*new_shared
;
4001 struct array_cache
**new_alien
= NULL
;
4003 for_each_online_node(node
) {
4005 if (use_alien_caches
) {
4006 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
4012 if (cachep
->shared
) {
4013 new_shared
= alloc_arraycache(node
,
4014 cachep
->shared
*cachep
->batchcount
,
4017 free_alien_cache(new_alien
);
4022 l3
= cachep
->nodelists
[node
];
4024 struct array_cache
*shared
= l3
->shared
;
4026 spin_lock_irq(&l3
->list_lock
);
4029 free_block(cachep
, shared
->entry
,
4030 shared
->avail
, node
);
4032 l3
->shared
= new_shared
;
4034 l3
->alien
= new_alien
;
4037 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4038 cachep
->batchcount
+ cachep
->num
;
4039 spin_unlock_irq(&l3
->list_lock
);
4041 free_alien_cache(new_alien
);
4044 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
4046 free_alien_cache(new_alien
);
4051 kmem_list3_init(l3
);
4052 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
4053 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
4054 l3
->shared
= new_shared
;
4055 l3
->alien
= new_alien
;
4056 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4057 cachep
->batchcount
+ cachep
->num
;
4058 cachep
->nodelists
[node
] = l3
;
4063 if (!cachep
->list
.next
) {
4064 /* Cache is not active yet. Roll back what we did */
4067 if (cachep
->nodelists
[node
]) {
4068 l3
= cachep
->nodelists
[node
];
4071 free_alien_cache(l3
->alien
);
4073 cachep
->nodelists
[node
] = NULL
;
4081 struct ccupdate_struct
{
4082 struct kmem_cache
*cachep
;
4083 struct array_cache
*new[0];
4086 static void do_ccupdate_local(void *info
)
4088 struct ccupdate_struct
*new = info
;
4089 struct array_cache
*old
;
4092 old
= cpu_cache_get(new->cachep
);
4094 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
4095 new->new[smp_processor_id()] = old
;
4098 /* Always called with the slab_mutex held */
4099 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
4100 int batchcount
, int shared
, gfp_t gfp
)
4102 struct ccupdate_struct
*new;
4105 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
4110 for_each_online_cpu(i
) {
4111 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
4114 for (i
--; i
>= 0; i
--)
4120 new->cachep
= cachep
;
4122 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
4125 cachep
->batchcount
= batchcount
;
4126 cachep
->limit
= limit
;
4127 cachep
->shared
= shared
;
4129 for_each_online_cpu(i
) {
4130 struct array_cache
*ccold
= new->new[i
];
4133 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4134 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4135 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4139 return alloc_kmemlist(cachep
, gfp
);
4142 /* Called with slab_mutex held always */
4143 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4149 * The head array serves three purposes:
4150 * - create a LIFO ordering, i.e. return objects that are cache-warm
4151 * - reduce the number of spinlock operations.
4152 * - reduce the number of linked list operations on the slab and
4153 * bufctl chains: array operations are cheaper.
4154 * The numbers are guessed, we should auto-tune as described by
4157 if (cachep
->size
> 131072)
4159 else if (cachep
->size
> PAGE_SIZE
)
4161 else if (cachep
->size
> 1024)
4163 else if (cachep
->size
> 256)
4169 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4170 * allocation behaviour: Most allocs on one cpu, most free operations
4171 * on another cpu. For these cases, an efficient object passing between
4172 * cpus is necessary. This is provided by a shared array. The array
4173 * replaces Bonwick's magazine layer.
4174 * On uniprocessor, it's functionally equivalent (but less efficient)
4175 * to a larger limit. Thus disabled by default.
4178 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4183 * With debugging enabled, large batchcount lead to excessively long
4184 * periods with disabled local interrupts. Limit the batchcount
4189 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4191 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4192 cachep
->name
, -err
);
4197 * Drain an array if it contains any elements taking the l3 lock only if
4198 * necessary. Note that the l3 listlock also protects the array_cache
4199 * if drain_array() is used on the shared array.
4201 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4202 struct array_cache
*ac
, int force
, int node
)
4206 if (!ac
|| !ac
->avail
)
4208 if (ac
->touched
&& !force
) {
4211 spin_lock_irq(&l3
->list_lock
);
4213 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4214 if (tofree
> ac
->avail
)
4215 tofree
= (ac
->avail
+ 1) / 2;
4216 free_block(cachep
, ac
->entry
, tofree
, node
);
4217 ac
->avail
-= tofree
;
4218 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4219 sizeof(void *) * ac
->avail
);
4221 spin_unlock_irq(&l3
->list_lock
);
4226 * cache_reap - Reclaim memory from caches.
4227 * @w: work descriptor
4229 * Called from workqueue/eventd every few seconds.
4231 * - clear the per-cpu caches for this CPU.
4232 * - return freeable pages to the main free memory pool.
4234 * If we cannot acquire the cache chain mutex then just give up - we'll try
4235 * again on the next iteration.
4237 static void cache_reap(struct work_struct
*w
)
4239 struct kmem_cache
*searchp
;
4240 struct kmem_list3
*l3
;
4241 int node
= numa_mem_id();
4242 struct delayed_work
*work
= to_delayed_work(w
);
4244 if (!mutex_trylock(&slab_mutex
))
4245 /* Give up. Setup the next iteration. */
4248 list_for_each_entry(searchp
, &slab_caches
, list
) {
4252 * We only take the l3 lock if absolutely necessary and we
4253 * have established with reasonable certainty that
4254 * we can do some work if the lock was obtained.
4256 l3
= searchp
->nodelists
[node
];
4258 reap_alien(searchp
, l3
);
4260 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4263 * These are racy checks but it does not matter
4264 * if we skip one check or scan twice.
4266 if (time_after(l3
->next_reap
, jiffies
))
4269 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4271 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4273 if (l3
->free_touched
)
4274 l3
->free_touched
= 0;
4278 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4279 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4280 STATS_ADD_REAPED(searchp
, freed
);
4286 mutex_unlock(&slab_mutex
);
4289 /* Set up the next iteration */
4290 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4293 #ifdef CONFIG_SLABINFO
4295 static void print_slabinfo_header(struct seq_file
*m
)
4298 * Output format version, so at least we can change it
4299 * without _too_ many complaints.
4302 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4304 seq_puts(m
, "slabinfo - version: 2.1\n");
4306 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4307 "<objperslab> <pagesperslab>");
4308 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4309 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4311 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4312 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4313 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4318 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4322 mutex_lock(&slab_mutex
);
4324 print_slabinfo_header(m
);
4326 return seq_list_start(&slab_caches
, *pos
);
4329 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4331 return seq_list_next(p
, &slab_caches
, pos
);
4334 static void s_stop(struct seq_file
*m
, void *p
)
4336 mutex_unlock(&slab_mutex
);
4339 static int s_show(struct seq_file
*m
, void *p
)
4341 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4343 unsigned long active_objs
;
4344 unsigned long num_objs
;
4345 unsigned long active_slabs
= 0;
4346 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4350 struct kmem_list3
*l3
;
4354 for_each_online_node(node
) {
4355 l3
= cachep
->nodelists
[node
];
4360 spin_lock_irq(&l3
->list_lock
);
4362 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4363 if (slabp
->inuse
!= cachep
->num
&& !error
)
4364 error
= "slabs_full accounting error";
4365 active_objs
+= cachep
->num
;
4368 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4369 if (slabp
->inuse
== cachep
->num
&& !error
)
4370 error
= "slabs_partial inuse accounting error";
4371 if (!slabp
->inuse
&& !error
)
4372 error
= "slabs_partial/inuse accounting error";
4373 active_objs
+= slabp
->inuse
;
4376 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4377 if (slabp
->inuse
&& !error
)
4378 error
= "slabs_free/inuse accounting error";
4381 free_objects
+= l3
->free_objects
;
4383 shared_avail
+= l3
->shared
->avail
;
4385 spin_unlock_irq(&l3
->list_lock
);
4387 num_slabs
+= active_slabs
;
4388 num_objs
= num_slabs
* cachep
->num
;
4389 if (num_objs
- active_objs
!= free_objects
&& !error
)
4390 error
= "free_objects accounting error";
4392 name
= cachep
->name
;
4394 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4396 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4397 name
, active_objs
, num_objs
, cachep
->size
,
4398 cachep
->num
, (1 << cachep
->gfporder
));
4399 seq_printf(m
, " : tunables %4u %4u %4u",
4400 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4401 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4402 active_slabs
, num_slabs
, shared_avail
);
4405 unsigned long high
= cachep
->high_mark
;
4406 unsigned long allocs
= cachep
->num_allocations
;
4407 unsigned long grown
= cachep
->grown
;
4408 unsigned long reaped
= cachep
->reaped
;
4409 unsigned long errors
= cachep
->errors
;
4410 unsigned long max_freeable
= cachep
->max_freeable
;
4411 unsigned long node_allocs
= cachep
->node_allocs
;
4412 unsigned long node_frees
= cachep
->node_frees
;
4413 unsigned long overflows
= cachep
->node_overflow
;
4415 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4416 "%4lu %4lu %4lu %4lu %4lu",
4417 allocs
, high
, grown
,
4418 reaped
, errors
, max_freeable
, node_allocs
,
4419 node_frees
, overflows
);
4423 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4424 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4425 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4426 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4428 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4429 allochit
, allocmiss
, freehit
, freemiss
);
4437 * slabinfo_op - iterator that generates /proc/slabinfo
4446 * num-pages-per-slab
4447 * + further values on SMP and with statistics enabled
4450 static const struct seq_operations slabinfo_op
= {
4457 #define MAX_SLABINFO_WRITE 128
4459 * slabinfo_write - Tuning for the slab allocator
4461 * @buffer: user buffer
4462 * @count: data length
4465 static ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4466 size_t count
, loff_t
*ppos
)
4468 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4469 int limit
, batchcount
, shared
, res
;
4470 struct kmem_cache
*cachep
;
4472 if (count
> MAX_SLABINFO_WRITE
)
4474 if (copy_from_user(&kbuf
, buffer
, count
))
4476 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4478 tmp
= strchr(kbuf
, ' ');
4483 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4486 /* Find the cache in the chain of caches. */
4487 mutex_lock(&slab_mutex
);
4489 list_for_each_entry(cachep
, &slab_caches
, list
) {
4490 if (!strcmp(cachep
->name
, kbuf
)) {
4491 if (limit
< 1 || batchcount
< 1 ||
4492 batchcount
> limit
|| shared
< 0) {
4495 res
= do_tune_cpucache(cachep
, limit
,
4502 mutex_unlock(&slab_mutex
);
4508 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4510 return seq_open(file
, &slabinfo_op
);
4513 static const struct file_operations proc_slabinfo_operations
= {
4514 .open
= slabinfo_open
,
4516 .write
= slabinfo_write
,
4517 .llseek
= seq_lseek
,
4518 .release
= seq_release
,
4521 #ifdef CONFIG_DEBUG_SLAB_LEAK
4523 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4525 mutex_lock(&slab_mutex
);
4526 return seq_list_start(&slab_caches
, *pos
);
4529 static inline int add_caller(unsigned long *n
, unsigned long v
)
4539 unsigned long *q
= p
+ 2 * i
;
4553 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4559 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4565 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4566 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4568 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4573 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4575 #ifdef CONFIG_KALLSYMS
4576 unsigned long offset
, size
;
4577 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4579 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4580 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4582 seq_printf(m
, " [%s]", modname
);
4586 seq_printf(m
, "%p", (void *)address
);
4589 static int leaks_show(struct seq_file
*m
, void *p
)
4591 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4593 struct kmem_list3
*l3
;
4595 unsigned long *n
= m
->private;
4599 if (!(cachep
->flags
& SLAB_STORE_USER
))
4601 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4604 /* OK, we can do it */
4608 for_each_online_node(node
) {
4609 l3
= cachep
->nodelists
[node
];
4614 spin_lock_irq(&l3
->list_lock
);
4616 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4617 handle_slab(n
, cachep
, slabp
);
4618 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4619 handle_slab(n
, cachep
, slabp
);
4620 spin_unlock_irq(&l3
->list_lock
);
4622 name
= cachep
->name
;
4624 /* Increase the buffer size */
4625 mutex_unlock(&slab_mutex
);
4626 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4628 /* Too bad, we are really out */
4630 mutex_lock(&slab_mutex
);
4633 *(unsigned long *)m
->private = n
[0] * 2;
4635 mutex_lock(&slab_mutex
);
4636 /* Now make sure this entry will be retried */
4640 for (i
= 0; i
< n
[1]; i
++) {
4641 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4642 show_symbol(m
, n
[2*i
+2]);
4649 static const struct seq_operations slabstats_op
= {
4650 .start
= leaks_start
,
4656 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4658 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4661 ret
= seq_open(file
, &slabstats_op
);
4663 struct seq_file
*m
= file
->private_data
;
4664 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4673 static const struct file_operations proc_slabstats_operations
= {
4674 .open
= slabstats_open
,
4676 .llseek
= seq_lseek
,
4677 .release
= seq_release_private
,
4681 static int __init
slab_proc_init(void)
4683 proc_create("slabinfo",S_IWUSR
|S_IRUSR
,NULL
,&proc_slabinfo_operations
);
4684 #ifdef CONFIG_DEBUG_SLAB_LEAK
4685 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4689 module_init(slab_proc_init
);
4693 * ksize - get the actual amount of memory allocated for a given object
4694 * @objp: Pointer to the object
4696 * kmalloc may internally round up allocations and return more memory
4697 * than requested. ksize() can be used to determine the actual amount of
4698 * memory allocated. The caller may use this additional memory, even though
4699 * a smaller amount of memory was initially specified with the kmalloc call.
4700 * The caller must guarantee that objp points to a valid object previously
4701 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4702 * must not be freed during the duration of the call.
4704 size_t ksize(const void *objp
)
4707 if (unlikely(objp
== ZERO_SIZE_PTR
))
4710 return virt_to_cache(objp
)->object_size
;
4712 EXPORT_SYMBOL(ksize
);