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;})
502 * Do not go above this order unless 0 objects fit into the slab or
503 * overridden on the command line.
505 #define SLAB_MAX_ORDER_HI 1
506 #define SLAB_MAX_ORDER_LO 0
507 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
508 static bool slab_max_order_set __initdata
;
510 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
512 struct page
*page
= virt_to_head_page(obj
);
513 return page
->slab_cache
;
516 static inline struct slab
*virt_to_slab(const void *obj
)
518 struct page
*page
= virt_to_head_page(obj
);
520 VM_BUG_ON(!PageSlab(page
));
521 return page
->slab_page
;
524 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
527 return slab
->s_mem
+ cache
->size
* idx
;
531 * We want to avoid an expensive divide : (offset / cache->size)
532 * Using the fact that size is a constant for a particular cache,
533 * we can replace (offset / cache->size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
537 const struct slab
*slab
, void *obj
)
539 u32 offset
= (obj
- slab
->s_mem
);
540 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 struct cache_sizes malloc_sizes
[] = {
547 #define CACHE(x) { .cs_size = (x) },
548 #include <linux/kmalloc_sizes.h>
552 EXPORT_SYMBOL(malloc_sizes
);
554 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
560 static struct cache_names __initdata cache_names
[] = {
561 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562 #include <linux/kmalloc_sizes.h>
567 static struct arraycache_init initarray_cache __initdata
=
568 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
569 static struct arraycache_init initarray_generic
=
570 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
572 /* internal cache of cache description objs */
573 static struct kmem_list3
*cache_cache_nodelists
[MAX_NUMNODES
];
574 static struct kmem_cache cache_cache
= {
575 .nodelists
= cache_cache_nodelists
,
577 .limit
= BOOT_CPUCACHE_ENTRIES
,
579 .size
= sizeof(struct kmem_cache
),
580 .name
= "kmem_cache",
583 #define BAD_ALIEN_MAGIC 0x01020304ul
585 #ifdef CONFIG_LOCKDEP
588 * Slab sometimes uses the kmalloc slabs to store the slab headers
589 * for other slabs "off slab".
590 * The locking for this is tricky in that it nests within the locks
591 * of all other slabs in a few places; to deal with this special
592 * locking we put on-slab caches into a separate lock-class.
594 * We set lock class for alien array caches which are up during init.
595 * The lock annotation will be lost if all cpus of a node goes down and
596 * then comes back up during hotplug
598 static struct lock_class_key on_slab_l3_key
;
599 static struct lock_class_key on_slab_alc_key
;
601 static struct lock_class_key debugobj_l3_key
;
602 static struct lock_class_key debugobj_alc_key
;
604 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
605 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
608 struct array_cache
**alc
;
609 struct kmem_list3
*l3
;
612 l3
= cachep
->nodelists
[q
];
616 lockdep_set_class(&l3
->list_lock
, l3_key
);
619 * FIXME: This check for BAD_ALIEN_MAGIC
620 * should go away when common slab code is taught to
621 * work even without alien caches.
622 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
623 * for alloc_alien_cache,
625 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
629 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
633 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
635 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
638 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
642 for_each_online_node(node
)
643 slab_set_debugobj_lock_classes_node(cachep
, node
);
646 static void init_node_lock_keys(int q
)
648 struct cache_sizes
*s
= malloc_sizes
;
653 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
654 struct kmem_list3
*l3
;
656 l3
= s
->cs_cachep
->nodelists
[q
];
657 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
660 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
661 &on_slab_alc_key
, q
);
665 static inline void init_lock_keys(void)
670 init_node_lock_keys(node
);
673 static void init_node_lock_keys(int q
)
677 static inline void init_lock_keys(void)
681 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
685 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
690 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
692 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
694 return cachep
->array
[smp_processor_id()];
697 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
700 struct cache_sizes
*csizep
= malloc_sizes
;
703 /* This happens if someone tries to call
704 * kmem_cache_create(), or __kmalloc(), before
705 * the generic caches are initialized.
707 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
710 return ZERO_SIZE_PTR
;
712 while (size
> csizep
->cs_size
)
716 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
717 * has cs_{dma,}cachep==NULL. Thus no special case
718 * for large kmalloc calls required.
720 #ifdef CONFIG_ZONE_DMA
721 if (unlikely(gfpflags
& GFP_DMA
))
722 return csizep
->cs_dmacachep
;
724 return csizep
->cs_cachep
;
727 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
729 return __find_general_cachep(size
, gfpflags
);
732 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
734 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
738 * Calculate the number of objects and left-over bytes for a given buffer size.
740 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
741 size_t align
, int flags
, size_t *left_over
,
746 size_t slab_size
= PAGE_SIZE
<< gfporder
;
749 * The slab management structure can be either off the slab or
750 * on it. For the latter case, the memory allocated for a
754 * - One kmem_bufctl_t for each object
755 * - Padding to respect alignment of @align
756 * - @buffer_size bytes for each object
758 * If the slab management structure is off the slab, then the
759 * alignment will already be calculated into the size. Because
760 * the slabs are all pages aligned, the objects will be at the
761 * correct alignment when allocated.
763 if (flags
& CFLGS_OFF_SLAB
) {
765 nr_objs
= slab_size
/ buffer_size
;
767 if (nr_objs
> SLAB_LIMIT
)
768 nr_objs
= SLAB_LIMIT
;
771 * Ignore padding for the initial guess. The padding
772 * is at most @align-1 bytes, and @buffer_size is at
773 * least @align. In the worst case, this result will
774 * be one greater than the number of objects that fit
775 * into the memory allocation when taking the padding
778 nr_objs
= (slab_size
- sizeof(struct slab
)) /
779 (buffer_size
+ sizeof(kmem_bufctl_t
));
782 * This calculated number will be either the right
783 * amount, or one greater than what we want.
785 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
789 if (nr_objs
> SLAB_LIMIT
)
790 nr_objs
= SLAB_LIMIT
;
792 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
795 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
798 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
800 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
803 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
804 function
, cachep
->name
, msg
);
806 add_taint(TAINT_BAD_PAGE
);
810 * By default on NUMA we use alien caches to stage the freeing of
811 * objects allocated from other nodes. This causes massive memory
812 * inefficiencies when using fake NUMA setup to split memory into a
813 * large number of small nodes, so it can be disabled on the command
817 static int use_alien_caches __read_mostly
= 1;
818 static int __init
noaliencache_setup(char *s
)
820 use_alien_caches
= 0;
823 __setup("noaliencache", noaliencache_setup
);
825 static int __init
slab_max_order_setup(char *str
)
827 get_option(&str
, &slab_max_order
);
828 slab_max_order
= slab_max_order
< 0 ? 0 :
829 min(slab_max_order
, MAX_ORDER
- 1);
830 slab_max_order_set
= true;
834 __setup("slab_max_order=", slab_max_order_setup
);
838 * Special reaping functions for NUMA systems called from cache_reap().
839 * These take care of doing round robin flushing of alien caches (containing
840 * objects freed on different nodes from which they were allocated) and the
841 * flushing of remote pcps by calling drain_node_pages.
843 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
845 static void init_reap_node(int cpu
)
849 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
850 if (node
== MAX_NUMNODES
)
851 node
= first_node(node_online_map
);
853 per_cpu(slab_reap_node
, cpu
) = node
;
856 static void next_reap_node(void)
858 int node
= __this_cpu_read(slab_reap_node
);
860 node
= next_node(node
, node_online_map
);
861 if (unlikely(node
>= MAX_NUMNODES
))
862 node
= first_node(node_online_map
);
863 __this_cpu_write(slab_reap_node
, node
);
867 #define init_reap_node(cpu) do { } while (0)
868 #define next_reap_node(void) do { } while (0)
872 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
873 * via the workqueue/eventd.
874 * Add the CPU number into the expiration time to minimize the possibility of
875 * the CPUs getting into lockstep and contending for the global cache chain
878 static void __cpuinit
start_cpu_timer(int cpu
)
880 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
883 * When this gets called from do_initcalls via cpucache_init(),
884 * init_workqueues() has already run, so keventd will be setup
887 if (keventd_up() && reap_work
->work
.func
== NULL
) {
889 INIT_DELAYED_WORK_DEFERRABLE(reap_work
, cache_reap
);
890 schedule_delayed_work_on(cpu
, reap_work
,
891 __round_jiffies_relative(HZ
, cpu
));
895 static struct array_cache
*alloc_arraycache(int node
, int entries
,
896 int batchcount
, gfp_t gfp
)
898 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
899 struct array_cache
*nc
= NULL
;
901 nc
= kmalloc_node(memsize
, gfp
, node
);
903 * The array_cache structures contain pointers to free object.
904 * However, when such objects are allocated or transferred to another
905 * cache the pointers are not cleared and they could be counted as
906 * valid references during a kmemleak scan. Therefore, kmemleak must
907 * not scan such objects.
909 kmemleak_no_scan(nc
);
913 nc
->batchcount
= batchcount
;
915 spin_lock_init(&nc
->lock
);
920 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
922 struct page
*page
= virt_to_page(slabp
->s_mem
);
924 return PageSlabPfmemalloc(page
);
927 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
928 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
929 struct array_cache
*ac
)
931 struct kmem_list3
*l3
= cachep
->nodelists
[numa_mem_id()];
935 if (!pfmemalloc_active
)
938 spin_lock_irqsave(&l3
->list_lock
, flags
);
939 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
940 if (is_slab_pfmemalloc(slabp
))
943 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
944 if (is_slab_pfmemalloc(slabp
))
947 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
948 if (is_slab_pfmemalloc(slabp
))
951 pfmemalloc_active
= false;
953 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
956 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
957 gfp_t flags
, bool force_refill
)
960 void *objp
= ac
->entry
[--ac
->avail
];
962 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
963 if (unlikely(is_obj_pfmemalloc(objp
))) {
964 struct kmem_list3
*l3
;
966 if (gfp_pfmemalloc_allowed(flags
)) {
967 clear_obj_pfmemalloc(&objp
);
971 /* The caller cannot use PFMEMALLOC objects, find another one */
972 for (i
= 1; i
< ac
->avail
; i
++) {
973 /* If a !PFMEMALLOC object is found, swap them */
974 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
976 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
977 ac
->entry
[ac
->avail
] = objp
;
983 * If there are empty slabs on the slabs_free list and we are
984 * being forced to refill the cache, mark this one !pfmemalloc.
986 l3
= cachep
->nodelists
[numa_mem_id()];
987 if (!list_empty(&l3
->slabs_free
) && force_refill
) {
988 struct slab
*slabp
= virt_to_slab(objp
);
989 ClearPageSlabPfmemalloc(virt_to_page(slabp
->s_mem
));
990 clear_obj_pfmemalloc(&objp
);
991 recheck_pfmemalloc_active(cachep
, ac
);
995 /* No !PFMEMALLOC objects available */
1003 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
1004 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
1008 if (unlikely(sk_memalloc_socks()))
1009 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
1011 objp
= ac
->entry
[--ac
->avail
];
1016 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1019 if (unlikely(pfmemalloc_active
)) {
1020 /* Some pfmemalloc slabs exist, check if this is one */
1021 struct page
*page
= virt_to_page(objp
);
1022 if (PageSlabPfmemalloc(page
))
1023 set_obj_pfmemalloc(&objp
);
1029 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1032 if (unlikely(sk_memalloc_socks()))
1033 objp
= __ac_put_obj(cachep
, ac
, objp
);
1035 ac
->entry
[ac
->avail
++] = objp
;
1039 * Transfer objects in one arraycache to another.
1040 * Locking must be handled by the caller.
1042 * Return the number of entries transferred.
1044 static int transfer_objects(struct array_cache
*to
,
1045 struct array_cache
*from
, unsigned int max
)
1047 /* Figure out how many entries to transfer */
1048 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
1053 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1054 sizeof(void *) *nr
);
1063 #define drain_alien_cache(cachep, alien) do { } while (0)
1064 #define reap_alien(cachep, l3) do { } while (0)
1066 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1068 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1071 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1075 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1080 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1086 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1087 gfp_t flags
, int nodeid
)
1092 #else /* CONFIG_NUMA */
1094 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1095 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1097 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1099 struct array_cache
**ac_ptr
;
1100 int memsize
= sizeof(void *) * nr_node_ids
;
1105 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1108 if (i
== node
|| !node_online(i
))
1110 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1112 for (i
--; i
>= 0; i
--)
1122 static void free_alien_cache(struct array_cache
**ac_ptr
)
1133 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1134 struct array_cache
*ac
, int node
)
1136 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1139 spin_lock(&rl3
->list_lock
);
1141 * Stuff objects into the remote nodes shared array first.
1142 * That way we could avoid the overhead of putting the objects
1143 * into the free lists and getting them back later.
1146 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1148 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1150 spin_unlock(&rl3
->list_lock
);
1155 * Called from cache_reap() to regularly drain alien caches round robin.
1157 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1159 int node
= __this_cpu_read(slab_reap_node
);
1162 struct array_cache
*ac
= l3
->alien
[node
];
1164 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1165 __drain_alien_cache(cachep
, ac
, node
);
1166 spin_unlock_irq(&ac
->lock
);
1171 static void drain_alien_cache(struct kmem_cache
*cachep
,
1172 struct array_cache
**alien
)
1175 struct array_cache
*ac
;
1176 unsigned long flags
;
1178 for_each_online_node(i
) {
1181 spin_lock_irqsave(&ac
->lock
, flags
);
1182 __drain_alien_cache(cachep
, ac
, i
);
1183 spin_unlock_irqrestore(&ac
->lock
, flags
);
1188 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1190 struct slab
*slabp
= virt_to_slab(objp
);
1191 int nodeid
= slabp
->nodeid
;
1192 struct kmem_list3
*l3
;
1193 struct array_cache
*alien
= NULL
;
1196 node
= numa_mem_id();
1199 * Make sure we are not freeing a object from another node to the array
1200 * cache on this cpu.
1202 if (likely(slabp
->nodeid
== node
))
1205 l3
= cachep
->nodelists
[node
];
1206 STATS_INC_NODEFREES(cachep
);
1207 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1208 alien
= l3
->alien
[nodeid
];
1209 spin_lock(&alien
->lock
);
1210 if (unlikely(alien
->avail
== alien
->limit
)) {
1211 STATS_INC_ACOVERFLOW(cachep
);
1212 __drain_alien_cache(cachep
, alien
, nodeid
);
1214 ac_put_obj(cachep
, alien
, objp
);
1215 spin_unlock(&alien
->lock
);
1217 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1218 free_block(cachep
, &objp
, 1, nodeid
);
1219 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1226 * Allocates and initializes nodelists for a node on each slab cache, used for
1227 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1228 * will be allocated off-node since memory is not yet online for the new node.
1229 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1232 * Must hold slab_mutex.
1234 static int init_cache_nodelists_node(int node
)
1236 struct kmem_cache
*cachep
;
1237 struct kmem_list3
*l3
;
1238 const int memsize
= sizeof(struct kmem_list3
);
1240 list_for_each_entry(cachep
, &slab_caches
, list
) {
1242 * Set up the size64 kmemlist for cpu before we can
1243 * begin anything. Make sure some other cpu on this
1244 * node has not already allocated this
1246 if (!cachep
->nodelists
[node
]) {
1247 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1250 kmem_list3_init(l3
);
1251 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1252 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1255 * The l3s don't come and go as CPUs come and
1256 * go. slab_mutex is sufficient
1259 cachep
->nodelists
[node
] = l3
;
1262 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1263 cachep
->nodelists
[node
]->free_limit
=
1264 (1 + nr_cpus_node(node
)) *
1265 cachep
->batchcount
+ cachep
->num
;
1266 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1271 static void __cpuinit
cpuup_canceled(long cpu
)
1273 struct kmem_cache
*cachep
;
1274 struct kmem_list3
*l3
= NULL
;
1275 int node
= cpu_to_mem(cpu
);
1276 const struct cpumask
*mask
= cpumask_of_node(node
);
1278 list_for_each_entry(cachep
, &slab_caches
, list
) {
1279 struct array_cache
*nc
;
1280 struct array_cache
*shared
;
1281 struct array_cache
**alien
;
1283 /* cpu is dead; no one can alloc from it. */
1284 nc
= cachep
->array
[cpu
];
1285 cachep
->array
[cpu
] = NULL
;
1286 l3
= cachep
->nodelists
[node
];
1289 goto free_array_cache
;
1291 spin_lock_irq(&l3
->list_lock
);
1293 /* Free limit for this kmem_list3 */
1294 l3
->free_limit
-= cachep
->batchcount
;
1296 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1298 if (!cpumask_empty(mask
)) {
1299 spin_unlock_irq(&l3
->list_lock
);
1300 goto free_array_cache
;
1303 shared
= l3
->shared
;
1305 free_block(cachep
, shared
->entry
,
1306 shared
->avail
, node
);
1313 spin_unlock_irq(&l3
->list_lock
);
1317 drain_alien_cache(cachep
, alien
);
1318 free_alien_cache(alien
);
1324 * In the previous loop, all the objects were freed to
1325 * the respective cache's slabs, now we can go ahead and
1326 * shrink each nodelist to its limit.
1328 list_for_each_entry(cachep
, &slab_caches
, list
) {
1329 l3
= cachep
->nodelists
[node
];
1332 drain_freelist(cachep
, l3
, l3
->free_objects
);
1336 static int __cpuinit
cpuup_prepare(long cpu
)
1338 struct kmem_cache
*cachep
;
1339 struct kmem_list3
*l3
= NULL
;
1340 int node
= cpu_to_mem(cpu
);
1344 * We need to do this right in the beginning since
1345 * alloc_arraycache's are going to use this list.
1346 * kmalloc_node allows us to add the slab to the right
1347 * kmem_list3 and not this cpu's kmem_list3
1349 err
= init_cache_nodelists_node(node
);
1354 * Now we can go ahead with allocating the shared arrays and
1357 list_for_each_entry(cachep
, &slab_caches
, list
) {
1358 struct array_cache
*nc
;
1359 struct array_cache
*shared
= NULL
;
1360 struct array_cache
**alien
= NULL
;
1362 nc
= alloc_arraycache(node
, cachep
->limit
,
1363 cachep
->batchcount
, GFP_KERNEL
);
1366 if (cachep
->shared
) {
1367 shared
= alloc_arraycache(node
,
1368 cachep
->shared
* cachep
->batchcount
,
1369 0xbaadf00d, GFP_KERNEL
);
1375 if (use_alien_caches
) {
1376 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1383 cachep
->array
[cpu
] = nc
;
1384 l3
= cachep
->nodelists
[node
];
1387 spin_lock_irq(&l3
->list_lock
);
1390 * We are serialised from CPU_DEAD or
1391 * CPU_UP_CANCELLED by the cpucontrol lock
1393 l3
->shared
= shared
;
1402 spin_unlock_irq(&l3
->list_lock
);
1404 free_alien_cache(alien
);
1405 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1406 slab_set_debugobj_lock_classes_node(cachep
, node
);
1408 init_node_lock_keys(node
);
1412 cpuup_canceled(cpu
);
1416 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1417 unsigned long action
, void *hcpu
)
1419 long cpu
= (long)hcpu
;
1423 case CPU_UP_PREPARE
:
1424 case CPU_UP_PREPARE_FROZEN
:
1425 mutex_lock(&slab_mutex
);
1426 err
= cpuup_prepare(cpu
);
1427 mutex_unlock(&slab_mutex
);
1430 case CPU_ONLINE_FROZEN
:
1431 start_cpu_timer(cpu
);
1433 #ifdef CONFIG_HOTPLUG_CPU
1434 case CPU_DOWN_PREPARE
:
1435 case CPU_DOWN_PREPARE_FROZEN
:
1437 * Shutdown cache reaper. Note that the slab_mutex is
1438 * held so that if cache_reap() is invoked it cannot do
1439 * anything expensive but will only modify reap_work
1440 * and reschedule the timer.
1442 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1443 /* Now the cache_reaper is guaranteed to be not running. */
1444 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1446 case CPU_DOWN_FAILED
:
1447 case CPU_DOWN_FAILED_FROZEN
:
1448 start_cpu_timer(cpu
);
1451 case CPU_DEAD_FROZEN
:
1453 * Even if all the cpus of a node are down, we don't free the
1454 * kmem_list3 of any cache. This to avoid a race between
1455 * cpu_down, and a kmalloc allocation from another cpu for
1456 * memory from the node of the cpu going down. The list3
1457 * structure is usually allocated from kmem_cache_create() and
1458 * gets destroyed at kmem_cache_destroy().
1462 case CPU_UP_CANCELED
:
1463 case CPU_UP_CANCELED_FROZEN
:
1464 mutex_lock(&slab_mutex
);
1465 cpuup_canceled(cpu
);
1466 mutex_unlock(&slab_mutex
);
1469 return notifier_from_errno(err
);
1472 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1473 &cpuup_callback
, NULL
, 0
1476 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1478 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1479 * Returns -EBUSY if all objects cannot be drained so that the node is not
1482 * Must hold slab_mutex.
1484 static int __meminit
drain_cache_nodelists_node(int node
)
1486 struct kmem_cache
*cachep
;
1489 list_for_each_entry(cachep
, &slab_caches
, list
) {
1490 struct kmem_list3
*l3
;
1492 l3
= cachep
->nodelists
[node
];
1496 drain_freelist(cachep
, l3
, l3
->free_objects
);
1498 if (!list_empty(&l3
->slabs_full
) ||
1499 !list_empty(&l3
->slabs_partial
)) {
1507 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1508 unsigned long action
, void *arg
)
1510 struct memory_notify
*mnb
= arg
;
1514 nid
= mnb
->status_change_nid
;
1519 case MEM_GOING_ONLINE
:
1520 mutex_lock(&slab_mutex
);
1521 ret
= init_cache_nodelists_node(nid
);
1522 mutex_unlock(&slab_mutex
);
1524 case MEM_GOING_OFFLINE
:
1525 mutex_lock(&slab_mutex
);
1526 ret
= drain_cache_nodelists_node(nid
);
1527 mutex_unlock(&slab_mutex
);
1531 case MEM_CANCEL_ONLINE
:
1532 case MEM_CANCEL_OFFLINE
:
1536 return notifier_from_errno(ret
);
1538 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1541 * swap the static kmem_list3 with kmalloced memory
1543 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1546 struct kmem_list3
*ptr
;
1548 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1551 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1553 * Do not assume that spinlocks can be initialized via memcpy:
1555 spin_lock_init(&ptr
->list_lock
);
1557 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1558 cachep
->nodelists
[nodeid
] = ptr
;
1562 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1563 * size of kmem_list3.
1565 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1569 for_each_online_node(node
) {
1570 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1571 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1573 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1578 * Initialisation. Called after the page allocator have been initialised and
1579 * before smp_init().
1581 void __init
kmem_cache_init(void)
1584 struct cache_sizes
*sizes
;
1585 struct cache_names
*names
;
1590 if (num_possible_nodes() == 1)
1591 use_alien_caches
= 0;
1593 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1594 kmem_list3_init(&initkmem_list3
[i
]);
1595 if (i
< MAX_NUMNODES
)
1596 cache_cache
.nodelists
[i
] = NULL
;
1598 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1601 * Fragmentation resistance on low memory - only use bigger
1602 * page orders on machines with more than 32MB of memory if
1603 * not overridden on the command line.
1605 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1606 slab_max_order
= SLAB_MAX_ORDER_HI
;
1608 /* Bootstrap is tricky, because several objects are allocated
1609 * from caches that do not exist yet:
1610 * 1) initialize the cache_cache cache: it contains the struct
1611 * kmem_cache structures of all caches, except cache_cache itself:
1612 * cache_cache is statically allocated.
1613 * Initially an __init data area is used for the head array and the
1614 * kmem_list3 structures, it's replaced with a kmalloc allocated
1615 * array at the end of the bootstrap.
1616 * 2) Create the first kmalloc cache.
1617 * The struct kmem_cache for the new cache is allocated normally.
1618 * An __init data area is used for the head array.
1619 * 3) Create the remaining kmalloc caches, with minimally sized
1621 * 4) Replace the __init data head arrays for cache_cache and the first
1622 * kmalloc cache with kmalloc allocated arrays.
1623 * 5) Replace the __init data for kmem_list3 for cache_cache and
1624 * the other cache's with kmalloc allocated memory.
1625 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1628 node
= numa_mem_id();
1630 /* 1) create the cache_cache */
1631 INIT_LIST_HEAD(&slab_caches
);
1632 list_add(&cache_cache
.list
, &slab_caches
);
1633 cache_cache
.colour_off
= cache_line_size();
1634 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1635 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1638 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1640 cache_cache
.size
= offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1641 nr_node_ids
* sizeof(struct kmem_list3
*);
1642 cache_cache
.object_size
= cache_cache
.size
;
1643 cache_cache
.size
= ALIGN(cache_cache
.size
,
1645 cache_cache
.reciprocal_buffer_size
=
1646 reciprocal_value(cache_cache
.size
);
1648 for (order
= 0; order
< MAX_ORDER
; order
++) {
1649 cache_estimate(order
, cache_cache
.size
,
1650 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1651 if (cache_cache
.num
)
1654 BUG_ON(!cache_cache
.num
);
1655 cache_cache
.gfporder
= order
;
1656 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1657 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1658 sizeof(struct slab
), cache_line_size());
1660 /* 2+3) create the kmalloc caches */
1661 sizes
= malloc_sizes
;
1662 names
= cache_names
;
1665 * Initialize the caches that provide memory for the array cache and the
1666 * kmem_list3 structures first. Without this, further allocations will
1670 sizes
[INDEX_AC
].cs_cachep
= __kmem_cache_create(names
[INDEX_AC
].name
,
1671 sizes
[INDEX_AC
].cs_size
,
1672 ARCH_KMALLOC_MINALIGN
,
1673 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1676 if (INDEX_AC
!= INDEX_L3
) {
1677 sizes
[INDEX_L3
].cs_cachep
=
1678 __kmem_cache_create(names
[INDEX_L3
].name
,
1679 sizes
[INDEX_L3
].cs_size
,
1680 ARCH_KMALLOC_MINALIGN
,
1681 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1685 slab_early_init
= 0;
1687 while (sizes
->cs_size
!= ULONG_MAX
) {
1689 * For performance, all the general caches are L1 aligned.
1690 * This should be particularly beneficial on SMP boxes, as it
1691 * eliminates "false sharing".
1692 * Note for systems short on memory removing the alignment will
1693 * allow tighter packing of the smaller caches.
1695 if (!sizes
->cs_cachep
) {
1696 sizes
->cs_cachep
= __kmem_cache_create(names
->name
,
1698 ARCH_KMALLOC_MINALIGN
,
1699 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1702 #ifdef CONFIG_ZONE_DMA
1703 sizes
->cs_dmacachep
= __kmem_cache_create(
1706 ARCH_KMALLOC_MINALIGN
,
1707 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1714 /* 4) Replace the bootstrap head arrays */
1716 struct array_cache
*ptr
;
1718 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1720 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1721 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1722 sizeof(struct arraycache_init
));
1724 * Do not assume that spinlocks can be initialized via memcpy:
1726 spin_lock_init(&ptr
->lock
);
1728 cache_cache
.array
[smp_processor_id()] = ptr
;
1730 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1732 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1733 != &initarray_generic
.cache
);
1734 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1735 sizeof(struct arraycache_init
));
1737 * Do not assume that spinlocks can be initialized via memcpy:
1739 spin_lock_init(&ptr
->lock
);
1741 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1744 /* 5) Replace the bootstrap kmem_list3's */
1748 for_each_online_node(nid
) {
1749 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1751 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1752 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1754 if (INDEX_AC
!= INDEX_L3
) {
1755 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1756 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1764 void __init
kmem_cache_init_late(void)
1766 struct kmem_cache
*cachep
;
1770 /* 6) resize the head arrays to their final sizes */
1771 mutex_lock(&slab_mutex
);
1772 list_for_each_entry(cachep
, &slab_caches
, list
)
1773 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1775 mutex_unlock(&slab_mutex
);
1777 /* Annotate slab for lockdep -- annotate the malloc caches */
1784 * Register a cpu startup notifier callback that initializes
1785 * cpu_cache_get for all new cpus
1787 register_cpu_notifier(&cpucache_notifier
);
1791 * Register a memory hotplug callback that initializes and frees
1794 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1798 * The reap timers are started later, with a module init call: That part
1799 * of the kernel is not yet operational.
1803 static int __init
cpucache_init(void)
1808 * Register the timers that return unneeded pages to the page allocator
1810 for_each_online_cpu(cpu
)
1811 start_cpu_timer(cpu
);
1817 __initcall(cpucache_init
);
1819 static noinline
void
1820 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1822 struct kmem_list3
*l3
;
1824 unsigned long flags
;
1828 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1830 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1831 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1833 for_each_online_node(node
) {
1834 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1835 unsigned long active_slabs
= 0, num_slabs
= 0;
1837 l3
= cachep
->nodelists
[node
];
1841 spin_lock_irqsave(&l3
->list_lock
, flags
);
1842 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
1843 active_objs
+= cachep
->num
;
1846 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
1847 active_objs
+= slabp
->inuse
;
1850 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
1853 free_objects
+= l3
->free_objects
;
1854 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
1856 num_slabs
+= active_slabs
;
1857 num_objs
= num_slabs
* cachep
->num
;
1859 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1860 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1866 * Interface to system's page allocator. No need to hold the cache-lock.
1868 * If we requested dmaable memory, we will get it. Even if we
1869 * did not request dmaable memory, we might get it, but that
1870 * would be relatively rare and ignorable.
1872 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1880 * Nommu uses slab's for process anonymous memory allocations, and thus
1881 * requires __GFP_COMP to properly refcount higher order allocations
1883 flags
|= __GFP_COMP
;
1886 flags
|= cachep
->allocflags
;
1887 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1888 flags
|= __GFP_RECLAIMABLE
;
1890 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1892 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1893 slab_out_of_memory(cachep
, flags
, nodeid
);
1897 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1898 if (unlikely(page
->pfmemalloc
))
1899 pfmemalloc_active
= true;
1901 nr_pages
= (1 << cachep
->gfporder
);
1902 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1903 add_zone_page_state(page_zone(page
),
1904 NR_SLAB_RECLAIMABLE
, nr_pages
);
1906 add_zone_page_state(page_zone(page
),
1907 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1908 for (i
= 0; i
< nr_pages
; i
++) {
1909 __SetPageSlab(page
+ i
);
1911 if (page
->pfmemalloc
)
1912 SetPageSlabPfmemalloc(page
+ i
);
1915 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1916 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1919 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1921 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1924 return page_address(page
);
1928 * Interface to system's page release.
1930 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1932 unsigned long i
= (1 << cachep
->gfporder
);
1933 struct page
*page
= virt_to_page(addr
);
1934 const unsigned long nr_freed
= i
;
1936 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1938 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1939 sub_zone_page_state(page_zone(page
),
1940 NR_SLAB_RECLAIMABLE
, nr_freed
);
1942 sub_zone_page_state(page_zone(page
),
1943 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1945 BUG_ON(!PageSlab(page
));
1946 __ClearPageSlabPfmemalloc(page
);
1947 __ClearPageSlab(page
);
1950 if (current
->reclaim_state
)
1951 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1952 free_pages((unsigned long)addr
, cachep
->gfporder
);
1955 static void kmem_rcu_free(struct rcu_head
*head
)
1957 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1958 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1960 kmem_freepages(cachep
, slab_rcu
->addr
);
1961 if (OFF_SLAB(cachep
))
1962 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1967 #ifdef CONFIG_DEBUG_PAGEALLOC
1968 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1969 unsigned long caller
)
1971 int size
= cachep
->object_size
;
1973 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1975 if (size
< 5 * sizeof(unsigned long))
1978 *addr
++ = 0x12345678;
1980 *addr
++ = smp_processor_id();
1981 size
-= 3 * sizeof(unsigned long);
1983 unsigned long *sptr
= &caller
;
1984 unsigned long svalue
;
1986 while (!kstack_end(sptr
)) {
1988 if (kernel_text_address(svalue
)) {
1990 size
-= sizeof(unsigned long);
1991 if (size
<= sizeof(unsigned long))
1997 *addr
++ = 0x87654321;
2001 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
2003 int size
= cachep
->object_size
;
2004 addr
= &((char *)addr
)[obj_offset(cachep
)];
2006 memset(addr
, val
, size
);
2007 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
2010 static void dump_line(char *data
, int offset
, int limit
)
2013 unsigned char error
= 0;
2016 printk(KERN_ERR
"%03x: ", offset
);
2017 for (i
= 0; i
< limit
; i
++) {
2018 if (data
[offset
+ i
] != POISON_FREE
) {
2019 error
= data
[offset
+ i
];
2023 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
2024 &data
[offset
], limit
, 1);
2026 if (bad_count
== 1) {
2027 error
^= POISON_FREE
;
2028 if (!(error
& (error
- 1))) {
2029 printk(KERN_ERR
"Single bit error detected. Probably "
2032 printk(KERN_ERR
"Run memtest86+ or a similar memory "
2035 printk(KERN_ERR
"Run a memory test tool.\n");
2044 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
2049 if (cachep
->flags
& SLAB_RED_ZONE
) {
2050 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
2051 *dbg_redzone1(cachep
, objp
),
2052 *dbg_redzone2(cachep
, objp
));
2055 if (cachep
->flags
& SLAB_STORE_USER
) {
2056 printk(KERN_ERR
"Last user: [<%p>]",
2057 *dbg_userword(cachep
, objp
));
2058 print_symbol("(%s)",
2059 (unsigned long)*dbg_userword(cachep
, objp
));
2062 realobj
= (char *)objp
+ obj_offset(cachep
);
2063 size
= cachep
->object_size
;
2064 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
2067 if (i
+ limit
> size
)
2069 dump_line(realobj
, i
, limit
);
2073 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
2079 realobj
= (char *)objp
+ obj_offset(cachep
);
2080 size
= cachep
->object_size
;
2082 for (i
= 0; i
< size
; i
++) {
2083 char exp
= POISON_FREE
;
2086 if (realobj
[i
] != exp
) {
2092 "Slab corruption (%s): %s start=%p, len=%d\n",
2093 print_tainted(), cachep
->name
, realobj
, size
);
2094 print_objinfo(cachep
, objp
, 0);
2096 /* Hexdump the affected line */
2099 if (i
+ limit
> size
)
2101 dump_line(realobj
, i
, limit
);
2104 /* Limit to 5 lines */
2110 /* Print some data about the neighboring objects, if they
2113 struct slab
*slabp
= virt_to_slab(objp
);
2116 objnr
= obj_to_index(cachep
, slabp
, objp
);
2118 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
2119 realobj
= (char *)objp
+ obj_offset(cachep
);
2120 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
2122 print_objinfo(cachep
, objp
, 2);
2124 if (objnr
+ 1 < cachep
->num
) {
2125 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
2126 realobj
= (char *)objp
+ obj_offset(cachep
);
2127 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
2129 print_objinfo(cachep
, objp
, 2);
2136 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2139 for (i
= 0; i
< cachep
->num
; i
++) {
2140 void *objp
= index_to_obj(cachep
, slabp
, i
);
2142 if (cachep
->flags
& SLAB_POISON
) {
2143 #ifdef CONFIG_DEBUG_PAGEALLOC
2144 if (cachep
->size
% PAGE_SIZE
== 0 &&
2146 kernel_map_pages(virt_to_page(objp
),
2147 cachep
->size
/ PAGE_SIZE
, 1);
2149 check_poison_obj(cachep
, objp
);
2151 check_poison_obj(cachep
, objp
);
2154 if (cachep
->flags
& SLAB_RED_ZONE
) {
2155 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2156 slab_error(cachep
, "start of a freed object "
2158 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2159 slab_error(cachep
, "end of a freed object "
2165 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2171 * slab_destroy - destroy and release all objects in a slab
2172 * @cachep: cache pointer being destroyed
2173 * @slabp: slab pointer being destroyed
2175 * Destroy all the objs in a slab, and release the mem back to the system.
2176 * Before calling the slab must have been unlinked from the cache. The
2177 * cache-lock is not held/needed.
2179 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2181 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2183 slab_destroy_debugcheck(cachep
, slabp
);
2184 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2185 struct slab_rcu
*slab_rcu
;
2187 slab_rcu
= (struct slab_rcu
*)slabp
;
2188 slab_rcu
->cachep
= cachep
;
2189 slab_rcu
->addr
= addr
;
2190 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2192 kmem_freepages(cachep
, addr
);
2193 if (OFF_SLAB(cachep
))
2194 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2198 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
2201 struct kmem_list3
*l3
;
2203 for_each_online_cpu(i
)
2204 kfree(cachep
->array
[i
]);
2206 /* NUMA: free the list3 structures */
2207 for_each_online_node(i
) {
2208 l3
= cachep
->nodelists
[i
];
2211 free_alien_cache(l3
->alien
);
2215 kmem_cache_free(&cache_cache
, cachep
);
2220 * calculate_slab_order - calculate size (page order) of slabs
2221 * @cachep: pointer to the cache that is being created
2222 * @size: size of objects to be created in this cache.
2223 * @align: required alignment for the objects.
2224 * @flags: slab allocation flags
2226 * Also calculates the number of objects per slab.
2228 * This could be made much more intelligent. For now, try to avoid using
2229 * high order pages for slabs. When the gfp() functions are more friendly
2230 * towards high-order requests, this should be changed.
2232 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2233 size_t size
, size_t align
, unsigned long flags
)
2235 unsigned long offslab_limit
;
2236 size_t left_over
= 0;
2239 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2243 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2247 if (flags
& CFLGS_OFF_SLAB
) {
2249 * Max number of objs-per-slab for caches which
2250 * use off-slab slabs. Needed to avoid a possible
2251 * looping condition in cache_grow().
2253 offslab_limit
= size
- sizeof(struct slab
);
2254 offslab_limit
/= sizeof(kmem_bufctl_t
);
2256 if (num
> offslab_limit
)
2260 /* Found something acceptable - save it away */
2262 cachep
->gfporder
= gfporder
;
2263 left_over
= remainder
;
2266 * A VFS-reclaimable slab tends to have most allocations
2267 * as GFP_NOFS and we really don't want to have to be allocating
2268 * higher-order pages when we are unable to shrink dcache.
2270 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2274 * Large number of objects is good, but very large slabs are
2275 * currently bad for the gfp()s.
2277 if (gfporder
>= slab_max_order
)
2281 * Acceptable internal fragmentation?
2283 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2289 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2291 if (slab_state
>= FULL
)
2292 return enable_cpucache(cachep
, gfp
);
2294 if (slab_state
== DOWN
) {
2296 * Note: the first kmem_cache_create must create the cache
2297 * that's used by kmalloc(24), otherwise the creation of
2298 * further caches will BUG().
2300 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2303 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2304 * the first cache, then we need to set up all its list3s,
2305 * otherwise the creation of further caches will BUG().
2307 set_up_list3s(cachep
, SIZE_AC
);
2308 if (INDEX_AC
== INDEX_L3
)
2309 slab_state
= PARTIAL_L3
;
2311 slab_state
= PARTIAL_ARRAYCACHE
;
2313 cachep
->array
[smp_processor_id()] =
2314 kmalloc(sizeof(struct arraycache_init
), gfp
);
2316 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2317 set_up_list3s(cachep
, SIZE_L3
);
2318 slab_state
= PARTIAL_L3
;
2321 for_each_online_node(node
) {
2322 cachep
->nodelists
[node
] =
2323 kmalloc_node(sizeof(struct kmem_list3
),
2325 BUG_ON(!cachep
->nodelists
[node
]);
2326 kmem_list3_init(cachep
->nodelists
[node
]);
2330 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2331 jiffies
+ REAPTIMEOUT_LIST3
+
2332 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2334 cpu_cache_get(cachep
)->avail
= 0;
2335 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2336 cpu_cache_get(cachep
)->batchcount
= 1;
2337 cpu_cache_get(cachep
)->touched
= 0;
2338 cachep
->batchcount
= 1;
2339 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2344 * __kmem_cache_create - Create a cache.
2345 * @name: A string which is used in /proc/slabinfo to identify this cache.
2346 * @size: The size of objects to be created in this cache.
2347 * @align: The required alignment for the objects.
2348 * @flags: SLAB flags
2349 * @ctor: A constructor for the objects.
2351 * Returns a ptr to the cache on success, NULL on failure.
2352 * Cannot be called within a int, but can be interrupted.
2353 * The @ctor is run when new pages are allocated by the cache.
2355 * @name must be valid until the cache is destroyed. This implies that
2356 * the module calling this has to destroy the cache before getting unloaded.
2360 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2361 * to catch references to uninitialised memory.
2363 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2364 * for buffer overruns.
2366 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2367 * cacheline. This can be beneficial if you're counting cycles as closely
2371 __kmem_cache_create (const char *name
, size_t size
, size_t align
,
2372 unsigned long flags
, void (*ctor
)(void *))
2374 size_t left_over
, slab_size
, ralign
;
2375 struct kmem_cache
*cachep
= NULL
;
2381 * Enable redzoning and last user accounting, except for caches with
2382 * large objects, if the increased size would increase the object size
2383 * above the next power of two: caches with object sizes just above a
2384 * power of two have a significant amount of internal fragmentation.
2386 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2387 2 * sizeof(unsigned long long)))
2388 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2389 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2390 flags
|= SLAB_POISON
;
2392 if (flags
& SLAB_DESTROY_BY_RCU
)
2393 BUG_ON(flags
& SLAB_POISON
);
2396 * Always checks flags, a caller might be expecting debug support which
2399 BUG_ON(flags
& ~CREATE_MASK
);
2402 * Check that size is in terms of words. This is needed to avoid
2403 * unaligned accesses for some archs when redzoning is used, and makes
2404 * sure any on-slab bufctl's are also correctly aligned.
2406 if (size
& (BYTES_PER_WORD
- 1)) {
2407 size
+= (BYTES_PER_WORD
- 1);
2408 size
&= ~(BYTES_PER_WORD
- 1);
2411 /* calculate the final buffer alignment: */
2413 /* 1) arch recommendation: can be overridden for debug */
2414 if (flags
& SLAB_HWCACHE_ALIGN
) {
2416 * Default alignment: as specified by the arch code. Except if
2417 * an object is really small, then squeeze multiple objects into
2420 ralign
= cache_line_size();
2421 while (size
<= ralign
/ 2)
2424 ralign
= BYTES_PER_WORD
;
2428 * Redzoning and user store require word alignment or possibly larger.
2429 * Note this will be overridden by architecture or caller mandated
2430 * alignment if either is greater than BYTES_PER_WORD.
2432 if (flags
& SLAB_STORE_USER
)
2433 ralign
= BYTES_PER_WORD
;
2435 if (flags
& SLAB_RED_ZONE
) {
2436 ralign
= REDZONE_ALIGN
;
2437 /* If redzoning, ensure that the second redzone is suitably
2438 * aligned, by adjusting the object size accordingly. */
2439 size
+= REDZONE_ALIGN
- 1;
2440 size
&= ~(REDZONE_ALIGN
- 1);
2443 /* 2) arch mandated alignment */
2444 if (ralign
< ARCH_SLAB_MINALIGN
) {
2445 ralign
= ARCH_SLAB_MINALIGN
;
2447 /* 3) caller mandated alignment */
2448 if (ralign
< align
) {
2451 /* disable debug if necessary */
2452 if (ralign
> __alignof__(unsigned long long))
2453 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2459 if (slab_is_available())
2464 /* Get cache's description obj. */
2465 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2469 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
2470 cachep
->object_size
= size
;
2471 cachep
->align
= align
;
2475 * Both debugging options require word-alignment which is calculated
2478 if (flags
& SLAB_RED_ZONE
) {
2479 /* add space for red zone words */
2480 cachep
->obj_offset
+= sizeof(unsigned long long);
2481 size
+= 2 * sizeof(unsigned long long);
2483 if (flags
& SLAB_STORE_USER
) {
2484 /* user store requires one word storage behind the end of
2485 * the real object. But if the second red zone needs to be
2486 * aligned to 64 bits, we must allow that much space.
2488 if (flags
& SLAB_RED_ZONE
)
2489 size
+= REDZONE_ALIGN
;
2491 size
+= BYTES_PER_WORD
;
2493 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2494 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2495 && cachep
->object_size
> cache_line_size()
2496 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2497 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2504 * Determine if the slab management is 'on' or 'off' slab.
2505 * (bootstrapping cannot cope with offslab caches so don't do
2506 * it too early on. Always use on-slab management when
2507 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2509 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2510 !(flags
& SLAB_NOLEAKTRACE
))
2512 * Size is large, assume best to place the slab management obj
2513 * off-slab (should allow better packing of objs).
2515 flags
|= CFLGS_OFF_SLAB
;
2517 size
= ALIGN(size
, align
);
2519 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2523 "kmem_cache_create: couldn't create cache %s.\n", name
);
2524 kmem_cache_free(&cache_cache
, cachep
);
2527 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2528 + sizeof(struct slab
), align
);
2531 * If the slab has been placed off-slab, and we have enough space then
2532 * move it on-slab. This is at the expense of any extra colouring.
2534 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2535 flags
&= ~CFLGS_OFF_SLAB
;
2536 left_over
-= slab_size
;
2539 if (flags
& CFLGS_OFF_SLAB
) {
2540 /* really off slab. No need for manual alignment */
2542 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2544 #ifdef CONFIG_PAGE_POISONING
2545 /* If we're going to use the generic kernel_map_pages()
2546 * poisoning, then it's going to smash the contents of
2547 * the redzone and userword anyhow, so switch them off.
2549 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2550 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2554 cachep
->colour_off
= cache_line_size();
2555 /* Offset must be a multiple of the alignment. */
2556 if (cachep
->colour_off
< align
)
2557 cachep
->colour_off
= align
;
2558 cachep
->colour
= left_over
/ cachep
->colour_off
;
2559 cachep
->slab_size
= slab_size
;
2560 cachep
->flags
= flags
;
2561 cachep
->allocflags
= 0;
2562 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2563 cachep
->allocflags
|= GFP_DMA
;
2564 cachep
->size
= size
;
2565 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2567 if (flags
& CFLGS_OFF_SLAB
) {
2568 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2570 * This is a possibility for one of the malloc_sizes caches.
2571 * But since we go off slab only for object size greater than
2572 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2573 * this should not happen at all.
2574 * But leave a BUG_ON for some lucky dude.
2576 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2578 cachep
->ctor
= ctor
;
2579 cachep
->name
= name
;
2581 if (setup_cpu_cache(cachep
, gfp
)) {
2582 __kmem_cache_destroy(cachep
);
2586 if (flags
& SLAB_DEBUG_OBJECTS
) {
2588 * Would deadlock through slab_destroy()->call_rcu()->
2589 * debug_object_activate()->kmem_cache_alloc().
2591 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2593 slab_set_debugobj_lock_classes(cachep
);
2596 /* cache setup completed, link it into the list */
2597 list_add(&cachep
->list
, &slab_caches
);
2602 static void check_irq_off(void)
2604 BUG_ON(!irqs_disabled());
2607 static void check_irq_on(void)
2609 BUG_ON(irqs_disabled());
2612 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2616 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2620 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2624 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2629 #define check_irq_off() do { } while(0)
2630 #define check_irq_on() do { } while(0)
2631 #define check_spinlock_acquired(x) do { } while(0)
2632 #define check_spinlock_acquired_node(x, y) do { } while(0)
2635 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2636 struct array_cache
*ac
,
2637 int force
, int node
);
2639 static void do_drain(void *arg
)
2641 struct kmem_cache
*cachep
= arg
;
2642 struct array_cache
*ac
;
2643 int node
= numa_mem_id();
2646 ac
= cpu_cache_get(cachep
);
2647 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2648 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2649 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2653 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2655 struct kmem_list3
*l3
;
2658 on_each_cpu(do_drain
, cachep
, 1);
2660 for_each_online_node(node
) {
2661 l3
= cachep
->nodelists
[node
];
2662 if (l3
&& l3
->alien
)
2663 drain_alien_cache(cachep
, l3
->alien
);
2666 for_each_online_node(node
) {
2667 l3
= cachep
->nodelists
[node
];
2669 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2674 * Remove slabs from the list of free slabs.
2675 * Specify the number of slabs to drain in tofree.
2677 * Returns the actual number of slabs released.
2679 static int drain_freelist(struct kmem_cache
*cache
,
2680 struct kmem_list3
*l3
, int tofree
)
2682 struct list_head
*p
;
2687 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2689 spin_lock_irq(&l3
->list_lock
);
2690 p
= l3
->slabs_free
.prev
;
2691 if (p
== &l3
->slabs_free
) {
2692 spin_unlock_irq(&l3
->list_lock
);
2696 slabp
= list_entry(p
, struct slab
, list
);
2698 BUG_ON(slabp
->inuse
);
2700 list_del(&slabp
->list
);
2702 * Safe to drop the lock. The slab is no longer linked
2705 l3
->free_objects
-= cache
->num
;
2706 spin_unlock_irq(&l3
->list_lock
);
2707 slab_destroy(cache
, slabp
);
2714 /* Called with slab_mutex held to protect against cpu hotplug */
2715 static int __cache_shrink(struct kmem_cache
*cachep
)
2718 struct kmem_list3
*l3
;
2720 drain_cpu_caches(cachep
);
2723 for_each_online_node(i
) {
2724 l3
= cachep
->nodelists
[i
];
2728 drain_freelist(cachep
, l3
, l3
->free_objects
);
2730 ret
+= !list_empty(&l3
->slabs_full
) ||
2731 !list_empty(&l3
->slabs_partial
);
2733 return (ret
? 1 : 0);
2737 * kmem_cache_shrink - Shrink a cache.
2738 * @cachep: The cache to shrink.
2740 * Releases as many slabs as possible for a cache.
2741 * To help debugging, a zero exit status indicates all slabs were released.
2743 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2746 BUG_ON(!cachep
|| in_interrupt());
2749 mutex_lock(&slab_mutex
);
2750 ret
= __cache_shrink(cachep
);
2751 mutex_unlock(&slab_mutex
);
2755 EXPORT_SYMBOL(kmem_cache_shrink
);
2758 * kmem_cache_destroy - delete a cache
2759 * @cachep: the cache to destroy
2761 * Remove a &struct kmem_cache object from the slab cache.
2763 * It is expected this function will be called by a module when it is
2764 * unloaded. This will remove the cache completely, and avoid a duplicate
2765 * cache being allocated each time a module is loaded and unloaded, if the
2766 * module doesn't have persistent in-kernel storage across loads and unloads.
2768 * The cache must be empty before calling this function.
2770 * The caller must guarantee that no one will allocate memory from the cache
2771 * during the kmem_cache_destroy().
2773 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2775 BUG_ON(!cachep
|| in_interrupt());
2777 /* Find the cache in the chain of caches. */
2779 mutex_lock(&slab_mutex
);
2781 * the chain is never empty, cache_cache is never destroyed
2783 list_del(&cachep
->list
);
2784 if (__cache_shrink(cachep
)) {
2785 slab_error(cachep
, "Can't free all objects");
2786 list_add(&cachep
->list
, &slab_caches
);
2787 mutex_unlock(&slab_mutex
);
2792 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2795 __kmem_cache_destroy(cachep
);
2796 mutex_unlock(&slab_mutex
);
2799 EXPORT_SYMBOL(kmem_cache_destroy
);
2802 * Get the memory for a slab management obj.
2803 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2804 * always come from malloc_sizes caches. The slab descriptor cannot
2805 * come from the same cache which is getting created because,
2806 * when we are searching for an appropriate cache for these
2807 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2808 * If we are creating a malloc_sizes cache here it would not be visible to
2809 * kmem_find_general_cachep till the initialization is complete.
2810 * Hence we cannot have slabp_cache same as the original cache.
2812 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2813 int colour_off
, gfp_t local_flags
,
2818 if (OFF_SLAB(cachep
)) {
2819 /* Slab management obj is off-slab. */
2820 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2821 local_flags
, nodeid
);
2823 * If the first object in the slab is leaked (it's allocated
2824 * but no one has a reference to it), we want to make sure
2825 * kmemleak does not treat the ->s_mem pointer as a reference
2826 * to the object. Otherwise we will not report the leak.
2828 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2833 slabp
= objp
+ colour_off
;
2834 colour_off
+= cachep
->slab_size
;
2837 slabp
->colouroff
= colour_off
;
2838 slabp
->s_mem
= objp
+ colour_off
;
2839 slabp
->nodeid
= nodeid
;
2844 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2846 return (kmem_bufctl_t
*) (slabp
+ 1);
2849 static void cache_init_objs(struct kmem_cache
*cachep
,
2854 for (i
= 0; i
< cachep
->num
; i
++) {
2855 void *objp
= index_to_obj(cachep
, slabp
, i
);
2857 /* need to poison the objs? */
2858 if (cachep
->flags
& SLAB_POISON
)
2859 poison_obj(cachep
, objp
, POISON_FREE
);
2860 if (cachep
->flags
& SLAB_STORE_USER
)
2861 *dbg_userword(cachep
, objp
) = NULL
;
2863 if (cachep
->flags
& SLAB_RED_ZONE
) {
2864 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2865 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2868 * Constructors are not allowed to allocate memory from the same
2869 * cache which they are a constructor for. Otherwise, deadlock.
2870 * They must also be threaded.
2872 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2873 cachep
->ctor(objp
+ obj_offset(cachep
));
2875 if (cachep
->flags
& SLAB_RED_ZONE
) {
2876 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2877 slab_error(cachep
, "constructor overwrote the"
2878 " end of an object");
2879 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2880 slab_error(cachep
, "constructor overwrote the"
2881 " start of an object");
2883 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2884 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2885 kernel_map_pages(virt_to_page(objp
),
2886 cachep
->size
/ PAGE_SIZE
, 0);
2891 slab_bufctl(slabp
)[i
] = i
+ 1;
2893 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2896 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2898 if (CONFIG_ZONE_DMA_FLAG
) {
2899 if (flags
& GFP_DMA
)
2900 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2902 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2906 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2909 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2913 next
= slab_bufctl(slabp
)[slabp
->free
];
2915 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2916 WARN_ON(slabp
->nodeid
!= nodeid
);
2923 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2924 void *objp
, int nodeid
)
2926 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2929 /* Verify that the slab belongs to the intended node */
2930 WARN_ON(slabp
->nodeid
!= nodeid
);
2932 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2933 printk(KERN_ERR
"slab: double free detected in cache "
2934 "'%s', objp %p\n", cachep
->name
, objp
);
2938 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2939 slabp
->free
= objnr
;
2944 * Map pages beginning at addr to the given cache and slab. This is required
2945 * for the slab allocator to be able to lookup the cache and slab of a
2946 * virtual address for kfree, ksize, and slab debugging.
2948 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2954 page
= virt_to_page(addr
);
2957 if (likely(!PageCompound(page
)))
2958 nr_pages
<<= cache
->gfporder
;
2961 page
->slab_cache
= cache
;
2962 page
->slab_page
= slab
;
2964 } while (--nr_pages
);
2968 * Grow (by 1) the number of slabs within a cache. This is called by
2969 * kmem_cache_alloc() when there are no active objs left in a cache.
2971 static int cache_grow(struct kmem_cache
*cachep
,
2972 gfp_t flags
, int nodeid
, void *objp
)
2977 struct kmem_list3
*l3
;
2980 * Be lazy and only check for valid flags here, keeping it out of the
2981 * critical path in kmem_cache_alloc().
2983 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2984 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2986 /* Take the l3 list lock to change the colour_next on this node */
2988 l3
= cachep
->nodelists
[nodeid
];
2989 spin_lock(&l3
->list_lock
);
2991 /* Get colour for the slab, and cal the next value. */
2992 offset
= l3
->colour_next
;
2994 if (l3
->colour_next
>= cachep
->colour
)
2995 l3
->colour_next
= 0;
2996 spin_unlock(&l3
->list_lock
);
2998 offset
*= cachep
->colour_off
;
3000 if (local_flags
& __GFP_WAIT
)
3004 * The test for missing atomic flag is performed here, rather than
3005 * the more obvious place, simply to reduce the critical path length
3006 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
3007 * will eventually be caught here (where it matters).
3009 kmem_flagcheck(cachep
, flags
);
3012 * Get mem for the objs. Attempt to allocate a physical page from
3016 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
3020 /* Get slab management. */
3021 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
3022 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
3026 slab_map_pages(cachep
, slabp
, objp
);
3028 cache_init_objs(cachep
, slabp
);
3030 if (local_flags
& __GFP_WAIT
)
3031 local_irq_disable();
3033 spin_lock(&l3
->list_lock
);
3035 /* Make slab active. */
3036 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
3037 STATS_INC_GROWN(cachep
);
3038 l3
->free_objects
+= cachep
->num
;
3039 spin_unlock(&l3
->list_lock
);
3042 kmem_freepages(cachep
, objp
);
3044 if (local_flags
& __GFP_WAIT
)
3045 local_irq_disable();
3052 * Perform extra freeing checks:
3053 * - detect bad pointers.
3054 * - POISON/RED_ZONE checking
3056 static void kfree_debugcheck(const void *objp
)
3058 if (!virt_addr_valid(objp
)) {
3059 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
3060 (unsigned long)objp
);
3065 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
3067 unsigned long long redzone1
, redzone2
;
3069 redzone1
= *dbg_redzone1(cache
, obj
);
3070 redzone2
= *dbg_redzone2(cache
, obj
);
3075 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
3078 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
3079 slab_error(cache
, "double free detected");
3081 slab_error(cache
, "memory outside object was overwritten");
3083 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3084 obj
, redzone1
, redzone2
);
3087 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
3088 unsigned long caller
)
3094 BUG_ON(virt_to_cache(objp
) != cachep
);
3096 objp
-= obj_offset(cachep
);
3097 kfree_debugcheck(objp
);
3098 page
= virt_to_head_page(objp
);
3100 slabp
= page
->slab_page
;
3102 if (cachep
->flags
& SLAB_RED_ZONE
) {
3103 verify_redzone_free(cachep
, objp
);
3104 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
3105 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
3107 if (cachep
->flags
& SLAB_STORE_USER
)
3108 *dbg_userword(cachep
, objp
) = (void *)caller
;
3110 objnr
= obj_to_index(cachep
, slabp
, objp
);
3112 BUG_ON(objnr
>= cachep
->num
);
3113 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
3115 #ifdef CONFIG_DEBUG_SLAB_LEAK
3116 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
3118 if (cachep
->flags
& SLAB_POISON
) {
3119 #ifdef CONFIG_DEBUG_PAGEALLOC
3120 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
3121 store_stackinfo(cachep
, objp
, caller
);
3122 kernel_map_pages(virt_to_page(objp
),
3123 cachep
->size
/ PAGE_SIZE
, 0);
3125 poison_obj(cachep
, objp
, POISON_FREE
);
3128 poison_obj(cachep
, objp
, POISON_FREE
);
3134 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
3139 /* Check slab's freelist to see if this obj is there. */
3140 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
3142 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3145 if (entries
!= cachep
->num
- slabp
->inuse
) {
3147 printk(KERN_ERR
"slab: Internal list corruption detected in "
3148 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3149 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
3151 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3152 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3158 #define kfree_debugcheck(x) do { } while(0)
3159 #define cache_free_debugcheck(x,objp,z) (objp)
3160 #define check_slabp(x,y) do { } while(0)
3163 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
3167 struct kmem_list3
*l3
;
3168 struct array_cache
*ac
;
3172 node
= numa_mem_id();
3173 if (unlikely(force_refill
))
3176 ac
= cpu_cache_get(cachep
);
3177 batchcount
= ac
->batchcount
;
3178 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3180 * If there was little recent activity on this cache, then
3181 * perform only a partial refill. Otherwise we could generate
3184 batchcount
= BATCHREFILL_LIMIT
;
3186 l3
= cachep
->nodelists
[node
];
3188 BUG_ON(ac
->avail
> 0 || !l3
);
3189 spin_lock(&l3
->list_lock
);
3191 /* See if we can refill from the shared array */
3192 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3193 l3
->shared
->touched
= 1;
3197 while (batchcount
> 0) {
3198 struct list_head
*entry
;
3200 /* Get slab alloc is to come from. */
3201 entry
= l3
->slabs_partial
.next
;
3202 if (entry
== &l3
->slabs_partial
) {
3203 l3
->free_touched
= 1;
3204 entry
= l3
->slabs_free
.next
;
3205 if (entry
== &l3
->slabs_free
)
3209 slabp
= list_entry(entry
, struct slab
, list
);
3210 check_slabp(cachep
, slabp
);
3211 check_spinlock_acquired(cachep
);
3214 * The slab was either on partial or free list so
3215 * there must be at least one object available for
3218 BUG_ON(slabp
->inuse
>= cachep
->num
);
3220 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3221 STATS_INC_ALLOCED(cachep
);
3222 STATS_INC_ACTIVE(cachep
);
3223 STATS_SET_HIGH(cachep
);
3225 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
3228 check_slabp(cachep
, slabp
);
3230 /* move slabp to correct slabp list: */
3231 list_del(&slabp
->list
);
3232 if (slabp
->free
== BUFCTL_END
)
3233 list_add(&slabp
->list
, &l3
->slabs_full
);
3235 list_add(&slabp
->list
, &l3
->slabs_partial
);
3239 l3
->free_objects
-= ac
->avail
;
3241 spin_unlock(&l3
->list_lock
);
3243 if (unlikely(!ac
->avail
)) {
3246 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3248 /* cache_grow can reenable interrupts, then ac could change. */
3249 ac
= cpu_cache_get(cachep
);
3251 /* no objects in sight? abort */
3252 if (!x
&& (ac
->avail
== 0 || force_refill
))
3255 if (!ac
->avail
) /* objects refilled by interrupt? */
3260 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
3263 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3266 might_sleep_if(flags
& __GFP_WAIT
);
3268 kmem_flagcheck(cachep
, flags
);
3273 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3274 gfp_t flags
, void *objp
, unsigned long caller
)
3278 if (cachep
->flags
& SLAB_POISON
) {
3279 #ifdef CONFIG_DEBUG_PAGEALLOC
3280 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3281 kernel_map_pages(virt_to_page(objp
),
3282 cachep
->size
/ PAGE_SIZE
, 1);
3284 check_poison_obj(cachep
, objp
);
3286 check_poison_obj(cachep
, objp
);
3288 poison_obj(cachep
, objp
, POISON_INUSE
);
3290 if (cachep
->flags
& SLAB_STORE_USER
)
3291 *dbg_userword(cachep
, objp
) = (void *)caller
;
3293 if (cachep
->flags
& SLAB_RED_ZONE
) {
3294 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3295 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3296 slab_error(cachep
, "double free, or memory outside"
3297 " object was overwritten");
3299 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3300 objp
, *dbg_redzone1(cachep
, objp
),
3301 *dbg_redzone2(cachep
, objp
));
3303 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3304 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3306 #ifdef CONFIG_DEBUG_SLAB_LEAK
3311 slabp
= virt_to_head_page(objp
)->slab_page
;
3312 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3313 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3316 objp
+= obj_offset(cachep
);
3317 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3319 if (ARCH_SLAB_MINALIGN
&&
3320 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3321 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3322 objp
, (int)ARCH_SLAB_MINALIGN
);
3327 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3330 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3332 if (cachep
== &cache_cache
)
3335 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3338 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3341 struct array_cache
*ac
;
3342 bool force_refill
= false;
3346 ac
= cpu_cache_get(cachep
);
3347 if (likely(ac
->avail
)) {
3349 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3352 * Allow for the possibility all avail objects are not allowed
3353 * by the current flags
3356 STATS_INC_ALLOCHIT(cachep
);
3359 force_refill
= true;
3362 STATS_INC_ALLOCMISS(cachep
);
3363 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3365 * the 'ac' may be updated by cache_alloc_refill(),
3366 * and kmemleak_erase() requires its correct value.
3368 ac
= cpu_cache_get(cachep
);
3372 * To avoid a false negative, if an object that is in one of the
3373 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3374 * treat the array pointers as a reference to the object.
3377 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3383 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3385 * If we are in_interrupt, then process context, including cpusets and
3386 * mempolicy, may not apply and should not be used for allocation policy.
3388 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3390 int nid_alloc
, nid_here
;
3392 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3394 nid_alloc
= nid_here
= numa_mem_id();
3395 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3396 nid_alloc
= cpuset_slab_spread_node();
3397 else if (current
->mempolicy
)
3398 nid_alloc
= slab_node();
3399 if (nid_alloc
!= nid_here
)
3400 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3405 * Fallback function if there was no memory available and no objects on a
3406 * certain node and fall back is permitted. First we scan all the
3407 * available nodelists for available objects. If that fails then we
3408 * perform an allocation without specifying a node. This allows the page
3409 * allocator to do its reclaim / fallback magic. We then insert the
3410 * slab into the proper nodelist and then allocate from it.
3412 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3414 struct zonelist
*zonelist
;
3418 enum zone_type high_zoneidx
= gfp_zone(flags
);
3421 unsigned int cpuset_mems_cookie
;
3423 if (flags
& __GFP_THISNODE
)
3426 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3429 cpuset_mems_cookie
= get_mems_allowed();
3430 zonelist
= node_zonelist(slab_node(), flags
);
3434 * Look through allowed nodes for objects available
3435 * from existing per node queues.
3437 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3438 nid
= zone_to_nid(zone
);
3440 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3441 cache
->nodelists
[nid
] &&
3442 cache
->nodelists
[nid
]->free_objects
) {
3443 obj
= ____cache_alloc_node(cache
,
3444 flags
| GFP_THISNODE
, nid
);
3452 * This allocation will be performed within the constraints
3453 * of the current cpuset / memory policy requirements.
3454 * We may trigger various forms of reclaim on the allowed
3455 * set and go into memory reserves if necessary.
3457 if (local_flags
& __GFP_WAIT
)
3459 kmem_flagcheck(cache
, flags
);
3460 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3461 if (local_flags
& __GFP_WAIT
)
3462 local_irq_disable();
3465 * Insert into the appropriate per node queues
3467 nid
= page_to_nid(virt_to_page(obj
));
3468 if (cache_grow(cache
, flags
, nid
, obj
)) {
3469 obj
= ____cache_alloc_node(cache
,
3470 flags
| GFP_THISNODE
, nid
);
3473 * Another processor may allocate the
3474 * objects in the slab since we are
3475 * not holding any locks.
3479 /* cache_grow already freed obj */
3485 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3491 * A interface to enable slab creation on nodeid
3493 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3496 struct list_head
*entry
;
3498 struct kmem_list3
*l3
;
3502 l3
= cachep
->nodelists
[nodeid
];
3507 spin_lock(&l3
->list_lock
);
3508 entry
= l3
->slabs_partial
.next
;
3509 if (entry
== &l3
->slabs_partial
) {
3510 l3
->free_touched
= 1;
3511 entry
= l3
->slabs_free
.next
;
3512 if (entry
== &l3
->slabs_free
)
3516 slabp
= list_entry(entry
, struct slab
, list
);
3517 check_spinlock_acquired_node(cachep
, nodeid
);
3518 check_slabp(cachep
, slabp
);
3520 STATS_INC_NODEALLOCS(cachep
);
3521 STATS_INC_ACTIVE(cachep
);
3522 STATS_SET_HIGH(cachep
);
3524 BUG_ON(slabp
->inuse
== cachep
->num
);
3526 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3527 check_slabp(cachep
, slabp
);
3529 /* move slabp to correct slabp list: */
3530 list_del(&slabp
->list
);
3532 if (slabp
->free
== BUFCTL_END
)
3533 list_add(&slabp
->list
, &l3
->slabs_full
);
3535 list_add(&slabp
->list
, &l3
->slabs_partial
);
3537 spin_unlock(&l3
->list_lock
);
3541 spin_unlock(&l3
->list_lock
);
3542 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3546 return fallback_alloc(cachep
, flags
);
3553 * kmem_cache_alloc_node - Allocate an object on the specified node
3554 * @cachep: The cache to allocate from.
3555 * @flags: See kmalloc().
3556 * @nodeid: node number of the target node.
3557 * @caller: return address of caller, used for debug information
3559 * Identical to kmem_cache_alloc but it will allocate memory on the given
3560 * node, which can improve the performance for cpu bound structures.
3562 * Fallback to other node is possible if __GFP_THISNODE is not set.
3564 static __always_inline
void *
3565 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3566 unsigned long caller
)
3568 unsigned long save_flags
;
3570 int slab_node
= numa_mem_id();
3572 flags
&= gfp_allowed_mask
;
3574 lockdep_trace_alloc(flags
);
3576 if (slab_should_failslab(cachep
, flags
))
3579 cache_alloc_debugcheck_before(cachep
, flags
);
3580 local_irq_save(save_flags
);
3582 if (nodeid
== NUMA_NO_NODE
)
3585 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3586 /* Node not bootstrapped yet */
3587 ptr
= fallback_alloc(cachep
, flags
);
3591 if (nodeid
== slab_node
) {
3593 * Use the locally cached objects if possible.
3594 * However ____cache_alloc does not allow fallback
3595 * to other nodes. It may fail while we still have
3596 * objects on other nodes available.
3598 ptr
= ____cache_alloc(cachep
, flags
);
3602 /* ___cache_alloc_node can fall back to other nodes */
3603 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3605 local_irq_restore(save_flags
);
3606 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3607 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3611 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3613 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3614 memset(ptr
, 0, cachep
->object_size
);
3619 static __always_inline
void *
3620 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3624 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3625 objp
= alternate_node_alloc(cache
, flags
);
3629 objp
= ____cache_alloc(cache
, flags
);
3632 * We may just have run out of memory on the local node.
3633 * ____cache_alloc_node() knows how to locate memory on other nodes
3636 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3643 static __always_inline
void *
3644 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3646 return ____cache_alloc(cachep
, flags
);
3649 #endif /* CONFIG_NUMA */
3651 static __always_inline
void *
3652 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3654 unsigned long save_flags
;
3657 flags
&= gfp_allowed_mask
;
3659 lockdep_trace_alloc(flags
);
3661 if (slab_should_failslab(cachep
, flags
))
3664 cache_alloc_debugcheck_before(cachep
, flags
);
3665 local_irq_save(save_flags
);
3666 objp
= __do_cache_alloc(cachep
, flags
);
3667 local_irq_restore(save_flags
);
3668 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3669 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3674 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3676 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3677 memset(objp
, 0, cachep
->object_size
);
3683 * Caller needs to acquire correct kmem_list's list_lock
3685 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3689 struct kmem_list3
*l3
;
3691 for (i
= 0; i
< nr_objects
; i
++) {
3695 clear_obj_pfmemalloc(&objpp
[i
]);
3698 slabp
= virt_to_slab(objp
);
3699 l3
= cachep
->nodelists
[node
];
3700 list_del(&slabp
->list
);
3701 check_spinlock_acquired_node(cachep
, node
);
3702 check_slabp(cachep
, slabp
);
3703 slab_put_obj(cachep
, slabp
, objp
, node
);
3704 STATS_DEC_ACTIVE(cachep
);
3706 check_slabp(cachep
, slabp
);
3708 /* fixup slab chains */
3709 if (slabp
->inuse
== 0) {
3710 if (l3
->free_objects
> l3
->free_limit
) {
3711 l3
->free_objects
-= cachep
->num
;
3712 /* No need to drop any previously held
3713 * lock here, even if we have a off-slab slab
3714 * descriptor it is guaranteed to come from
3715 * a different cache, refer to comments before
3718 slab_destroy(cachep
, slabp
);
3720 list_add(&slabp
->list
, &l3
->slabs_free
);
3723 /* Unconditionally move a slab to the end of the
3724 * partial list on free - maximum time for the
3725 * other objects to be freed, too.
3727 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3732 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3735 struct kmem_list3
*l3
;
3736 int node
= numa_mem_id();
3738 batchcount
= ac
->batchcount
;
3740 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3743 l3
= cachep
->nodelists
[node
];
3744 spin_lock(&l3
->list_lock
);
3746 struct array_cache
*shared_array
= l3
->shared
;
3747 int max
= shared_array
->limit
- shared_array
->avail
;
3749 if (batchcount
> max
)
3751 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3752 ac
->entry
, sizeof(void *) * batchcount
);
3753 shared_array
->avail
+= batchcount
;
3758 free_block(cachep
, ac
->entry
, batchcount
, node
);
3763 struct list_head
*p
;
3765 p
= l3
->slabs_free
.next
;
3766 while (p
!= &(l3
->slabs_free
)) {
3769 slabp
= list_entry(p
, struct slab
, list
);
3770 BUG_ON(slabp
->inuse
);
3775 STATS_SET_FREEABLE(cachep
, i
);
3778 spin_unlock(&l3
->list_lock
);
3779 ac
->avail
-= batchcount
;
3780 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3784 * Release an obj back to its cache. If the obj has a constructed state, it must
3785 * be in this state _before_ it is released. Called with disabled ints.
3787 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3788 unsigned long caller
)
3790 struct array_cache
*ac
= cpu_cache_get(cachep
);
3793 kmemleak_free_recursive(objp
, cachep
->flags
);
3794 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3796 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3799 * Skip calling cache_free_alien() when the platform is not numa.
3800 * This will avoid cache misses that happen while accessing slabp (which
3801 * is per page memory reference) to get nodeid. Instead use a global
3802 * variable to skip the call, which is mostly likely to be present in
3805 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3808 if (likely(ac
->avail
< ac
->limit
)) {
3809 STATS_INC_FREEHIT(cachep
);
3811 STATS_INC_FREEMISS(cachep
);
3812 cache_flusharray(cachep
, ac
);
3815 ac_put_obj(cachep
, ac
, objp
);
3819 * kmem_cache_alloc - Allocate an object
3820 * @cachep: The cache to allocate from.
3821 * @flags: See kmalloc().
3823 * Allocate an object from this cache. The flags are only relevant
3824 * if the cache has no available objects.
3826 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3828 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3830 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3831 cachep
->object_size
, cachep
->size
, flags
);
3835 EXPORT_SYMBOL(kmem_cache_alloc
);
3837 #ifdef CONFIG_TRACING
3839 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3843 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3845 trace_kmalloc(_RET_IP_
, ret
,
3846 size
, cachep
->size
, flags
);
3849 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3853 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3855 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3857 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3858 cachep
->object_size
, cachep
->size
,
3863 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3865 #ifdef CONFIG_TRACING
3866 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3873 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP
);
3875 trace_kmalloc_node(_RET_IP_
, ret
,
3880 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3883 static __always_inline
void *
3884 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3886 struct kmem_cache
*cachep
;
3888 cachep
= kmem_find_general_cachep(size
, flags
);
3889 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3891 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3894 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3895 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3897 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3899 EXPORT_SYMBOL(__kmalloc_node
);
3901 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3902 int node
, unsigned long caller
)
3904 return __do_kmalloc_node(size
, flags
, node
, caller
);
3906 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3908 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3910 return __do_kmalloc_node(size
, flags
, node
, 0);
3912 EXPORT_SYMBOL(__kmalloc_node
);
3913 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3914 #endif /* CONFIG_NUMA */
3917 * __do_kmalloc - allocate memory
3918 * @size: how many bytes of memory are required.
3919 * @flags: the type of memory to allocate (see kmalloc).
3920 * @caller: function caller for debug tracking of the caller
3922 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3923 unsigned long caller
)
3925 struct kmem_cache
*cachep
;
3928 /* If you want to save a few bytes .text space: replace
3930 * Then kmalloc uses the uninlined functions instead of the inline
3933 cachep
= __find_general_cachep(size
, flags
);
3934 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3936 ret
= slab_alloc(cachep
, flags
, caller
);
3938 trace_kmalloc(caller
, ret
,
3939 size
, cachep
->size
, flags
);
3945 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3946 void *__kmalloc(size_t size
, gfp_t flags
)
3948 return __do_kmalloc(size
, flags
, _RET_IP_
);
3950 EXPORT_SYMBOL(__kmalloc
);
3952 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3954 return __do_kmalloc(size
, flags
, caller
);
3956 EXPORT_SYMBOL(__kmalloc_track_caller
);
3959 void *__kmalloc(size_t size
, gfp_t flags
)
3961 return __do_kmalloc(size
, flags
, 0);
3963 EXPORT_SYMBOL(__kmalloc
);
3967 * kmem_cache_free - Deallocate an object
3968 * @cachep: The cache the allocation was from.
3969 * @objp: The previously allocated object.
3971 * Free an object which was previously allocated from this
3974 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3976 unsigned long flags
;
3978 local_irq_save(flags
);
3979 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3980 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3981 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3982 __cache_free(cachep
, objp
, _RET_IP_
);
3983 local_irq_restore(flags
);
3985 trace_kmem_cache_free(_RET_IP_
, objp
);
3987 EXPORT_SYMBOL(kmem_cache_free
);
3990 * kfree - free previously allocated memory
3991 * @objp: pointer returned by kmalloc.
3993 * If @objp is NULL, no operation is performed.
3995 * Don't free memory not originally allocated by kmalloc()
3996 * or you will run into trouble.
3998 void kfree(const void *objp
)
4000 struct kmem_cache
*c
;
4001 unsigned long flags
;
4003 trace_kfree(_RET_IP_
, objp
);
4005 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
4007 local_irq_save(flags
);
4008 kfree_debugcheck(objp
);
4009 c
= virt_to_cache(objp
);
4010 debug_check_no_locks_freed(objp
, c
->object_size
);
4012 debug_check_no_obj_freed(objp
, c
->object_size
);
4013 __cache_free(c
, (void *)objp
, _RET_IP_
);
4014 local_irq_restore(flags
);
4016 EXPORT_SYMBOL(kfree
);
4018 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
4020 return cachep
->object_size
;
4022 EXPORT_SYMBOL(kmem_cache_size
);
4025 * This initializes kmem_list3 or resizes various caches for all nodes.
4027 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
4030 struct kmem_list3
*l3
;
4031 struct array_cache
*new_shared
;
4032 struct array_cache
**new_alien
= NULL
;
4034 for_each_online_node(node
) {
4036 if (use_alien_caches
) {
4037 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
4043 if (cachep
->shared
) {
4044 new_shared
= alloc_arraycache(node
,
4045 cachep
->shared
*cachep
->batchcount
,
4048 free_alien_cache(new_alien
);
4053 l3
= cachep
->nodelists
[node
];
4055 struct array_cache
*shared
= l3
->shared
;
4057 spin_lock_irq(&l3
->list_lock
);
4060 free_block(cachep
, shared
->entry
,
4061 shared
->avail
, node
);
4063 l3
->shared
= new_shared
;
4065 l3
->alien
= new_alien
;
4068 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4069 cachep
->batchcount
+ cachep
->num
;
4070 spin_unlock_irq(&l3
->list_lock
);
4072 free_alien_cache(new_alien
);
4075 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
4077 free_alien_cache(new_alien
);
4082 kmem_list3_init(l3
);
4083 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
4084 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
4085 l3
->shared
= new_shared
;
4086 l3
->alien
= new_alien
;
4087 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4088 cachep
->batchcount
+ cachep
->num
;
4089 cachep
->nodelists
[node
] = l3
;
4094 if (!cachep
->list
.next
) {
4095 /* Cache is not active yet. Roll back what we did */
4098 if (cachep
->nodelists
[node
]) {
4099 l3
= cachep
->nodelists
[node
];
4102 free_alien_cache(l3
->alien
);
4104 cachep
->nodelists
[node
] = NULL
;
4112 struct ccupdate_struct
{
4113 struct kmem_cache
*cachep
;
4114 struct array_cache
*new[0];
4117 static void do_ccupdate_local(void *info
)
4119 struct ccupdate_struct
*new = info
;
4120 struct array_cache
*old
;
4123 old
= cpu_cache_get(new->cachep
);
4125 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
4126 new->new[smp_processor_id()] = old
;
4129 /* Always called with the slab_mutex held */
4130 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
4131 int batchcount
, int shared
, gfp_t gfp
)
4133 struct ccupdate_struct
*new;
4136 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
4141 for_each_online_cpu(i
) {
4142 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
4145 for (i
--; i
>= 0; i
--)
4151 new->cachep
= cachep
;
4153 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
4156 cachep
->batchcount
= batchcount
;
4157 cachep
->limit
= limit
;
4158 cachep
->shared
= shared
;
4160 for_each_online_cpu(i
) {
4161 struct array_cache
*ccold
= new->new[i
];
4164 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4165 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4166 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4170 return alloc_kmemlist(cachep
, gfp
);
4173 /* Called with slab_mutex held always */
4174 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4180 * The head array serves three purposes:
4181 * - create a LIFO ordering, i.e. return objects that are cache-warm
4182 * - reduce the number of spinlock operations.
4183 * - reduce the number of linked list operations on the slab and
4184 * bufctl chains: array operations are cheaper.
4185 * The numbers are guessed, we should auto-tune as described by
4188 if (cachep
->size
> 131072)
4190 else if (cachep
->size
> PAGE_SIZE
)
4192 else if (cachep
->size
> 1024)
4194 else if (cachep
->size
> 256)
4200 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4201 * allocation behaviour: Most allocs on one cpu, most free operations
4202 * on another cpu. For these cases, an efficient object passing between
4203 * cpus is necessary. This is provided by a shared array. The array
4204 * replaces Bonwick's magazine layer.
4205 * On uniprocessor, it's functionally equivalent (but less efficient)
4206 * to a larger limit. Thus disabled by default.
4209 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4214 * With debugging enabled, large batchcount lead to excessively long
4215 * periods with disabled local interrupts. Limit the batchcount
4220 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4222 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4223 cachep
->name
, -err
);
4228 * Drain an array if it contains any elements taking the l3 lock only if
4229 * necessary. Note that the l3 listlock also protects the array_cache
4230 * if drain_array() is used on the shared array.
4232 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4233 struct array_cache
*ac
, int force
, int node
)
4237 if (!ac
|| !ac
->avail
)
4239 if (ac
->touched
&& !force
) {
4242 spin_lock_irq(&l3
->list_lock
);
4244 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4245 if (tofree
> ac
->avail
)
4246 tofree
= (ac
->avail
+ 1) / 2;
4247 free_block(cachep
, ac
->entry
, tofree
, node
);
4248 ac
->avail
-= tofree
;
4249 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4250 sizeof(void *) * ac
->avail
);
4252 spin_unlock_irq(&l3
->list_lock
);
4257 * cache_reap - Reclaim memory from caches.
4258 * @w: work descriptor
4260 * Called from workqueue/eventd every few seconds.
4262 * - clear the per-cpu caches for this CPU.
4263 * - return freeable pages to the main free memory pool.
4265 * If we cannot acquire the cache chain mutex then just give up - we'll try
4266 * again on the next iteration.
4268 static void cache_reap(struct work_struct
*w
)
4270 struct kmem_cache
*searchp
;
4271 struct kmem_list3
*l3
;
4272 int node
= numa_mem_id();
4273 struct delayed_work
*work
= to_delayed_work(w
);
4275 if (!mutex_trylock(&slab_mutex
))
4276 /* Give up. Setup the next iteration. */
4279 list_for_each_entry(searchp
, &slab_caches
, list
) {
4283 * We only take the l3 lock if absolutely necessary and we
4284 * have established with reasonable certainty that
4285 * we can do some work if the lock was obtained.
4287 l3
= searchp
->nodelists
[node
];
4289 reap_alien(searchp
, l3
);
4291 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4294 * These are racy checks but it does not matter
4295 * if we skip one check or scan twice.
4297 if (time_after(l3
->next_reap
, jiffies
))
4300 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4302 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4304 if (l3
->free_touched
)
4305 l3
->free_touched
= 0;
4309 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4310 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4311 STATS_ADD_REAPED(searchp
, freed
);
4317 mutex_unlock(&slab_mutex
);
4320 /* Set up the next iteration */
4321 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4324 #ifdef CONFIG_SLABINFO
4326 static void print_slabinfo_header(struct seq_file
*m
)
4329 * Output format version, so at least we can change it
4330 * without _too_ many complaints.
4333 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4335 seq_puts(m
, "slabinfo - version: 2.1\n");
4337 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4338 "<objperslab> <pagesperslab>");
4339 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4340 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4342 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4343 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4344 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4349 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4353 mutex_lock(&slab_mutex
);
4355 print_slabinfo_header(m
);
4357 return seq_list_start(&slab_caches
, *pos
);
4360 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4362 return seq_list_next(p
, &slab_caches
, pos
);
4365 static void s_stop(struct seq_file
*m
, void *p
)
4367 mutex_unlock(&slab_mutex
);
4370 static int s_show(struct seq_file
*m
, void *p
)
4372 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4374 unsigned long active_objs
;
4375 unsigned long num_objs
;
4376 unsigned long active_slabs
= 0;
4377 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4381 struct kmem_list3
*l3
;
4385 for_each_online_node(node
) {
4386 l3
= cachep
->nodelists
[node
];
4391 spin_lock_irq(&l3
->list_lock
);
4393 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4394 if (slabp
->inuse
!= cachep
->num
&& !error
)
4395 error
= "slabs_full accounting error";
4396 active_objs
+= cachep
->num
;
4399 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4400 if (slabp
->inuse
== cachep
->num
&& !error
)
4401 error
= "slabs_partial inuse accounting error";
4402 if (!slabp
->inuse
&& !error
)
4403 error
= "slabs_partial/inuse accounting error";
4404 active_objs
+= slabp
->inuse
;
4407 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4408 if (slabp
->inuse
&& !error
)
4409 error
= "slabs_free/inuse accounting error";
4412 free_objects
+= l3
->free_objects
;
4414 shared_avail
+= l3
->shared
->avail
;
4416 spin_unlock_irq(&l3
->list_lock
);
4418 num_slabs
+= active_slabs
;
4419 num_objs
= num_slabs
* cachep
->num
;
4420 if (num_objs
- active_objs
!= free_objects
&& !error
)
4421 error
= "free_objects accounting error";
4423 name
= cachep
->name
;
4425 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4427 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4428 name
, active_objs
, num_objs
, cachep
->size
,
4429 cachep
->num
, (1 << cachep
->gfporder
));
4430 seq_printf(m
, " : tunables %4u %4u %4u",
4431 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4432 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4433 active_slabs
, num_slabs
, shared_avail
);
4436 unsigned long high
= cachep
->high_mark
;
4437 unsigned long allocs
= cachep
->num_allocations
;
4438 unsigned long grown
= cachep
->grown
;
4439 unsigned long reaped
= cachep
->reaped
;
4440 unsigned long errors
= cachep
->errors
;
4441 unsigned long max_freeable
= cachep
->max_freeable
;
4442 unsigned long node_allocs
= cachep
->node_allocs
;
4443 unsigned long node_frees
= cachep
->node_frees
;
4444 unsigned long overflows
= cachep
->node_overflow
;
4446 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4447 "%4lu %4lu %4lu %4lu %4lu",
4448 allocs
, high
, grown
,
4449 reaped
, errors
, max_freeable
, node_allocs
,
4450 node_frees
, overflows
);
4454 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4455 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4456 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4457 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4459 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4460 allochit
, allocmiss
, freehit
, freemiss
);
4468 * slabinfo_op - iterator that generates /proc/slabinfo
4477 * num-pages-per-slab
4478 * + further values on SMP and with statistics enabled
4481 static const struct seq_operations slabinfo_op
= {
4488 #define MAX_SLABINFO_WRITE 128
4490 * slabinfo_write - Tuning for the slab allocator
4492 * @buffer: user buffer
4493 * @count: data length
4496 static ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4497 size_t count
, loff_t
*ppos
)
4499 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4500 int limit
, batchcount
, shared
, res
;
4501 struct kmem_cache
*cachep
;
4503 if (count
> MAX_SLABINFO_WRITE
)
4505 if (copy_from_user(&kbuf
, buffer
, count
))
4507 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4509 tmp
= strchr(kbuf
, ' ');
4514 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4517 /* Find the cache in the chain of caches. */
4518 mutex_lock(&slab_mutex
);
4520 list_for_each_entry(cachep
, &slab_caches
, list
) {
4521 if (!strcmp(cachep
->name
, kbuf
)) {
4522 if (limit
< 1 || batchcount
< 1 ||
4523 batchcount
> limit
|| shared
< 0) {
4526 res
= do_tune_cpucache(cachep
, limit
,
4533 mutex_unlock(&slab_mutex
);
4539 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4541 return seq_open(file
, &slabinfo_op
);
4544 static const struct file_operations proc_slabinfo_operations
= {
4545 .open
= slabinfo_open
,
4547 .write
= slabinfo_write
,
4548 .llseek
= seq_lseek
,
4549 .release
= seq_release
,
4552 #ifdef CONFIG_DEBUG_SLAB_LEAK
4554 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4556 mutex_lock(&slab_mutex
);
4557 return seq_list_start(&slab_caches
, *pos
);
4560 static inline int add_caller(unsigned long *n
, unsigned long v
)
4570 unsigned long *q
= p
+ 2 * i
;
4584 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4590 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4596 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4597 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4599 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4604 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4606 #ifdef CONFIG_KALLSYMS
4607 unsigned long offset
, size
;
4608 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4610 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4611 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4613 seq_printf(m
, " [%s]", modname
);
4617 seq_printf(m
, "%p", (void *)address
);
4620 static int leaks_show(struct seq_file
*m
, void *p
)
4622 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4624 struct kmem_list3
*l3
;
4626 unsigned long *n
= m
->private;
4630 if (!(cachep
->flags
& SLAB_STORE_USER
))
4632 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4635 /* OK, we can do it */
4639 for_each_online_node(node
) {
4640 l3
= cachep
->nodelists
[node
];
4645 spin_lock_irq(&l3
->list_lock
);
4647 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4648 handle_slab(n
, cachep
, slabp
);
4649 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4650 handle_slab(n
, cachep
, slabp
);
4651 spin_unlock_irq(&l3
->list_lock
);
4653 name
= cachep
->name
;
4655 /* Increase the buffer size */
4656 mutex_unlock(&slab_mutex
);
4657 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4659 /* Too bad, we are really out */
4661 mutex_lock(&slab_mutex
);
4664 *(unsigned long *)m
->private = n
[0] * 2;
4666 mutex_lock(&slab_mutex
);
4667 /* Now make sure this entry will be retried */
4671 for (i
= 0; i
< n
[1]; i
++) {
4672 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4673 show_symbol(m
, n
[2*i
+2]);
4680 static const struct seq_operations slabstats_op
= {
4681 .start
= leaks_start
,
4687 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4689 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4692 ret
= seq_open(file
, &slabstats_op
);
4694 struct seq_file
*m
= file
->private_data
;
4695 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4704 static const struct file_operations proc_slabstats_operations
= {
4705 .open
= slabstats_open
,
4707 .llseek
= seq_lseek
,
4708 .release
= seq_release_private
,
4712 static int __init
slab_proc_init(void)
4714 proc_create("slabinfo",S_IWUSR
|S_IRUSR
,NULL
,&proc_slabinfo_operations
);
4715 #ifdef CONFIG_DEBUG_SLAB_LEAK
4716 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4720 module_init(slab_proc_init
);
4724 * ksize - get the actual amount of memory allocated for a given object
4725 * @objp: Pointer to the object
4727 * kmalloc may internally round up allocations and return more memory
4728 * than requested. ksize() can be used to determine the actual amount of
4729 * memory allocated. The caller may use this additional memory, even though
4730 * a smaller amount of memory was initially specified with the kmalloc call.
4731 * The caller must guarantee that objp points to a valid object previously
4732 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4733 * must not be freed during the duration of the call.
4735 size_t ksize(const void *objp
)
4738 if (unlikely(objp
== ZERO_SIZE_PTR
))
4741 return virt_to_cache(objp
)->object_size
;
4743 EXPORT_SYMBOL(ksize
);