3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
121 #include <net/sock.h>
123 #include <asm/cacheflush.h>
124 #include <asm/tlbflush.h>
125 #include <asm/page.h>
127 #include <trace/events/kmem.h>
129 #include "internal.h"
132 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
133 * 0 for faster, smaller code (especially in the critical paths).
135 * STATS - 1 to collect stats for /proc/slabinfo.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 #ifdef CONFIG_DEBUG_SLAB
144 #define FORCED_DEBUG 1
148 #define FORCED_DEBUG 0
151 /* Shouldn't this be in a header file somewhere? */
152 #define BYTES_PER_WORD sizeof(void *)
153 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155 #ifndef ARCH_KMALLOC_FLAGS
156 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 * true if a page was allocated from pfmemalloc reserves for network-based
163 static bool pfmemalloc_active __read_mostly
;
165 /* Legal flag mask for kmem_cache_create(). */
167 # define CREATE_MASK (SLAB_RED_ZONE | \
168 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
171 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
172 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
173 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
175 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
179 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
185 * Bufctl's are used for linking objs within a slab
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 typedef unsigned int kmem_bufctl_t
;
202 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
204 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
205 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
210 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
211 * arrange for kmem_freepages to be called via RCU. This is useful if
212 * we need to approach a kernel structure obliquely, from its address
213 * obtained without the usual locking. We can lock the structure to
214 * stabilize it and check it's still at the given address, only if we
215 * can be sure that the memory has not been meanwhile reused for some
216 * other kind of object (which our subsystem's lock might corrupt).
218 * rcu_read_lock before reading the address, then rcu_read_unlock after
219 * taking the spinlock within the structure expected at that address.
222 struct rcu_head head
;
223 struct kmem_cache
*cachep
;
230 * Manages the objs in a slab. Placed either at the beginning of mem allocated
231 * for a slab, or allocated from an general cache.
232 * Slabs are chained into three list: fully used, partial, fully free slabs.
237 struct list_head list
;
238 unsigned long colouroff
;
239 void *s_mem
; /* including colour offset */
240 unsigned int inuse
; /* num of objs active in slab */
242 unsigned short nodeid
;
244 struct slab_rcu __slab_cover_slab_rcu
;
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
256 * The limit is stored in the per-cpu structure to reduce the data cache
263 unsigned int batchcount
;
264 unsigned int touched
;
267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
271 * Entries should not be directly dereferenced as
272 * entries belonging to slabs marked pfmemalloc will
273 * have the lower bits set SLAB_OBJ_PFMEMALLOC
277 #define SLAB_OBJ_PFMEMALLOC 1
278 static inline bool is_obj_pfmemalloc(void *objp
)
280 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
283 static inline void set_obj_pfmemalloc(void **objp
)
285 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
289 static inline void clear_obj_pfmemalloc(void **objp
)
291 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
295 * bootstrap: The caches do not work without cpuarrays anymore, but the
296 * cpuarrays are allocated from the generic caches...
298 #define BOOT_CPUCACHE_ENTRIES 1
299 struct arraycache_init
{
300 struct array_cache cache
;
301 void *entries
[BOOT_CPUCACHE_ENTRIES
];
305 * The slab lists for all objects.
308 struct list_head slabs_partial
; /* partial list first, better asm code */
309 struct list_head slabs_full
;
310 struct list_head slabs_free
;
311 unsigned long free_objects
;
312 unsigned int free_limit
;
313 unsigned int colour_next
; /* Per-node cache coloring */
314 spinlock_t list_lock
;
315 struct array_cache
*shared
; /* shared per node */
316 struct array_cache
**alien
; /* on other nodes */
317 unsigned long next_reap
; /* updated without locking */
318 int free_touched
; /* updated without locking */
322 * Need this for bootstrapping a per node allocator.
324 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
325 static struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
326 #define CACHE_CACHE 0
327 #define SIZE_AC MAX_NUMNODES
328 #define SIZE_L3 (2 * MAX_NUMNODES)
330 static int drain_freelist(struct kmem_cache
*cache
,
331 struct kmem_list3
*l3
, int tofree
);
332 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
334 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
335 static void cache_reap(struct work_struct
*unused
);
338 * This function must be completely optimized away if a constant is passed to
339 * it. Mostly the same as what is in linux/slab.h except it returns an index.
341 static __always_inline
int index_of(const size_t size
)
343 extern void __bad_size(void);
345 if (__builtin_constant_p(size
)) {
353 #include <linux/kmalloc_sizes.h>
361 static int slab_early_init
= 1;
363 #define INDEX_AC index_of(sizeof(struct arraycache_init))
364 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
366 static void kmem_list3_init(struct kmem_list3
*parent
)
368 INIT_LIST_HEAD(&parent
->slabs_full
);
369 INIT_LIST_HEAD(&parent
->slabs_partial
);
370 INIT_LIST_HEAD(&parent
->slabs_free
);
371 parent
->shared
= NULL
;
372 parent
->alien
= NULL
;
373 parent
->colour_next
= 0;
374 spin_lock_init(&parent
->list_lock
);
375 parent
->free_objects
= 0;
376 parent
->free_touched
= 0;
379 #define MAKE_LIST(cachep, listp, slab, nodeid) \
381 INIT_LIST_HEAD(listp); \
382 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
385 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
387 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
388 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
389 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
392 #define CFLGS_OFF_SLAB (0x80000000UL)
393 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
395 #define BATCHREFILL_LIMIT 16
397 * Optimization question: fewer reaps means less probability for unnessary
398 * cpucache drain/refill cycles.
400 * OTOH the cpuarrays can contain lots of objects,
401 * which could lock up otherwise freeable slabs.
403 #define REAPTIMEOUT_CPUC (2*HZ)
404 #define REAPTIMEOUT_LIST3 (4*HZ)
407 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
408 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
409 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
410 #define STATS_INC_GROWN(x) ((x)->grown++)
411 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
412 #define STATS_SET_HIGH(x) \
414 if ((x)->num_active > (x)->high_mark) \
415 (x)->high_mark = (x)->num_active; \
417 #define STATS_INC_ERR(x) ((x)->errors++)
418 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
419 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
420 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
421 #define STATS_SET_FREEABLE(x, i) \
423 if ((x)->max_freeable < i) \
424 (x)->max_freeable = i; \
426 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
427 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
428 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
429 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
431 #define STATS_INC_ACTIVE(x) do { } while (0)
432 #define STATS_DEC_ACTIVE(x) do { } while (0)
433 #define STATS_INC_ALLOCED(x) do { } while (0)
434 #define STATS_INC_GROWN(x) do { } while (0)
435 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
436 #define STATS_SET_HIGH(x) do { } while (0)
437 #define STATS_INC_ERR(x) do { } while (0)
438 #define STATS_INC_NODEALLOCS(x) do { } while (0)
439 #define STATS_INC_NODEFREES(x) do { } while (0)
440 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
441 #define STATS_SET_FREEABLE(x, i) do { } while (0)
442 #define STATS_INC_ALLOCHIT(x) do { } while (0)
443 #define STATS_INC_ALLOCMISS(x) do { } while (0)
444 #define STATS_INC_FREEHIT(x) do { } while (0)
445 #define STATS_INC_FREEMISS(x) do { } while (0)
451 * memory layout of objects:
453 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
454 * the end of an object is aligned with the end of the real
455 * allocation. Catches writes behind the end of the allocation.
456 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
458 * cachep->obj_offset: The real object.
459 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
460 * cachep->size - 1* BYTES_PER_WORD: last caller address
461 * [BYTES_PER_WORD long]
463 static int obj_offset(struct kmem_cache
*cachep
)
465 return cachep
->obj_offset
;
468 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
470 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
471 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
472 sizeof(unsigned long long));
475 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
477 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
478 if (cachep
->flags
& SLAB_STORE_USER
)
479 return (unsigned long long *)(objp
+ cachep
->size
-
480 sizeof(unsigned long long) -
482 return (unsigned long long *) (objp
+ cachep
->size
-
483 sizeof(unsigned long long));
486 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
488 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
489 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
494 #define obj_offset(x) 0
495 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
496 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
497 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
501 #ifdef CONFIG_TRACING
502 size_t slab_buffer_size(struct kmem_cache
*cachep
)
506 EXPORT_SYMBOL(slab_buffer_size
);
510 * Do not go above this order unless 0 objects fit into the slab or
511 * overridden on the command line.
513 #define SLAB_MAX_ORDER_HI 1
514 #define SLAB_MAX_ORDER_LO 0
515 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
516 static bool slab_max_order_set __initdata
;
518 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
520 page
= compound_head(page
);
521 BUG_ON(!PageSlab(page
));
522 return page
->slab_cache
;
525 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
527 struct page
*page
= virt_to_head_page(obj
);
528 return page
->slab_cache
;
531 static inline struct slab
*virt_to_slab(const void *obj
)
533 struct page
*page
= virt_to_head_page(obj
);
535 VM_BUG_ON(!PageSlab(page
));
536 return page
->slab_page
;
539 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
542 return slab
->s_mem
+ cache
->size
* idx
;
546 * We want to avoid an expensive divide : (offset / cache->size)
547 * Using the fact that size is a constant for a particular cache,
548 * we can replace (offset / cache->size) by
549 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
551 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
552 const struct slab
*slab
, void *obj
)
554 u32 offset
= (obj
- slab
->s_mem
);
555 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
559 * These are the default caches for kmalloc. Custom caches can have other sizes.
561 struct cache_sizes malloc_sizes
[] = {
562 #define CACHE(x) { .cs_size = (x) },
563 #include <linux/kmalloc_sizes.h>
567 EXPORT_SYMBOL(malloc_sizes
);
569 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
575 static struct cache_names __initdata cache_names
[] = {
576 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
577 #include <linux/kmalloc_sizes.h>
582 static struct arraycache_init initarray_cache __initdata
=
583 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
584 static struct arraycache_init initarray_generic
=
585 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
587 /* internal cache of cache description objs */
588 static struct kmem_list3
*cache_cache_nodelists
[MAX_NUMNODES
];
589 static struct kmem_cache cache_cache
= {
590 .nodelists
= cache_cache_nodelists
,
592 .limit
= BOOT_CPUCACHE_ENTRIES
,
594 .size
= sizeof(struct kmem_cache
),
595 .name
= "kmem_cache",
598 #define BAD_ALIEN_MAGIC 0x01020304ul
600 #ifdef CONFIG_LOCKDEP
603 * Slab sometimes uses the kmalloc slabs to store the slab headers
604 * for other slabs "off slab".
605 * The locking for this is tricky in that it nests within the locks
606 * of all other slabs in a few places; to deal with this special
607 * locking we put on-slab caches into a separate lock-class.
609 * We set lock class for alien array caches which are up during init.
610 * The lock annotation will be lost if all cpus of a node goes down and
611 * then comes back up during hotplug
613 static struct lock_class_key on_slab_l3_key
;
614 static struct lock_class_key on_slab_alc_key
;
616 static struct lock_class_key debugobj_l3_key
;
617 static struct lock_class_key debugobj_alc_key
;
619 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
620 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
623 struct array_cache
**alc
;
624 struct kmem_list3
*l3
;
627 l3
= cachep
->nodelists
[q
];
631 lockdep_set_class(&l3
->list_lock
, l3_key
);
634 * FIXME: This check for BAD_ALIEN_MAGIC
635 * should go away when common slab code is taught to
636 * work even without alien caches.
637 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
638 * for alloc_alien_cache,
640 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
644 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
648 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
650 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
653 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
657 for_each_online_node(node
)
658 slab_set_debugobj_lock_classes_node(cachep
, node
);
661 static void init_node_lock_keys(int q
)
663 struct cache_sizes
*s
= malloc_sizes
;
668 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
669 struct kmem_list3
*l3
;
671 l3
= s
->cs_cachep
->nodelists
[q
];
672 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
675 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
676 &on_slab_alc_key
, q
);
680 static inline void init_lock_keys(void)
685 init_node_lock_keys(node
);
688 static void init_node_lock_keys(int q
)
692 static inline void init_lock_keys(void)
696 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
700 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
705 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
707 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
709 return cachep
->array
[smp_processor_id()];
712 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
715 struct cache_sizes
*csizep
= malloc_sizes
;
718 /* This happens if someone tries to call
719 * kmem_cache_create(), or __kmalloc(), before
720 * the generic caches are initialized.
722 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
725 return ZERO_SIZE_PTR
;
727 while (size
> csizep
->cs_size
)
731 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
732 * has cs_{dma,}cachep==NULL. Thus no special case
733 * for large kmalloc calls required.
735 #ifdef CONFIG_ZONE_DMA
736 if (unlikely(gfpflags
& GFP_DMA
))
737 return csizep
->cs_dmacachep
;
739 return csizep
->cs_cachep
;
742 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
744 return __find_general_cachep(size
, gfpflags
);
747 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
749 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
753 * Calculate the number of objects and left-over bytes for a given buffer size.
755 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
756 size_t align
, int flags
, size_t *left_over
,
761 size_t slab_size
= PAGE_SIZE
<< gfporder
;
764 * The slab management structure can be either off the slab or
765 * on it. For the latter case, the memory allocated for a
769 * - One kmem_bufctl_t for each object
770 * - Padding to respect alignment of @align
771 * - @buffer_size bytes for each object
773 * If the slab management structure is off the slab, then the
774 * alignment will already be calculated into the size. Because
775 * the slabs are all pages aligned, the objects will be at the
776 * correct alignment when allocated.
778 if (flags
& CFLGS_OFF_SLAB
) {
780 nr_objs
= slab_size
/ buffer_size
;
782 if (nr_objs
> SLAB_LIMIT
)
783 nr_objs
= SLAB_LIMIT
;
786 * Ignore padding for the initial guess. The padding
787 * is at most @align-1 bytes, and @buffer_size is at
788 * least @align. In the worst case, this result will
789 * be one greater than the number of objects that fit
790 * into the memory allocation when taking the padding
793 nr_objs
= (slab_size
- sizeof(struct slab
)) /
794 (buffer_size
+ sizeof(kmem_bufctl_t
));
797 * This calculated number will be either the right
798 * amount, or one greater than what we want.
800 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
804 if (nr_objs
> SLAB_LIMIT
)
805 nr_objs
= SLAB_LIMIT
;
807 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
810 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
813 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
815 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
818 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
819 function
, cachep
->name
, msg
);
824 * By default on NUMA we use alien caches to stage the freeing of
825 * objects allocated from other nodes. This causes massive memory
826 * inefficiencies when using fake NUMA setup to split memory into a
827 * large number of small nodes, so it can be disabled on the command
831 static int use_alien_caches __read_mostly
= 1;
832 static int __init
noaliencache_setup(char *s
)
834 use_alien_caches
= 0;
837 __setup("noaliencache", noaliencache_setup
);
839 static int __init
slab_max_order_setup(char *str
)
841 get_option(&str
, &slab_max_order
);
842 slab_max_order
= slab_max_order
< 0 ? 0 :
843 min(slab_max_order
, MAX_ORDER
- 1);
844 slab_max_order_set
= true;
848 __setup("slab_max_order=", slab_max_order_setup
);
852 * Special reaping functions for NUMA systems called from cache_reap().
853 * These take care of doing round robin flushing of alien caches (containing
854 * objects freed on different nodes from which they were allocated) and the
855 * flushing of remote pcps by calling drain_node_pages.
857 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
859 static void init_reap_node(int cpu
)
863 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
864 if (node
== MAX_NUMNODES
)
865 node
= first_node(node_online_map
);
867 per_cpu(slab_reap_node
, cpu
) = node
;
870 static void next_reap_node(void)
872 int node
= __this_cpu_read(slab_reap_node
);
874 node
= next_node(node
, node_online_map
);
875 if (unlikely(node
>= MAX_NUMNODES
))
876 node
= first_node(node_online_map
);
877 __this_cpu_write(slab_reap_node
, node
);
881 #define init_reap_node(cpu) do { } while (0)
882 #define next_reap_node(void) do { } while (0)
886 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
887 * via the workqueue/eventd.
888 * Add the CPU number into the expiration time to minimize the possibility of
889 * the CPUs getting into lockstep and contending for the global cache chain
892 static void __cpuinit
start_cpu_timer(int cpu
)
894 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
897 * When this gets called from do_initcalls via cpucache_init(),
898 * init_workqueues() has already run, so keventd will be setup
901 if (keventd_up() && reap_work
->work
.func
== NULL
) {
903 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
904 schedule_delayed_work_on(cpu
, reap_work
,
905 __round_jiffies_relative(HZ
, cpu
));
909 static struct array_cache
*alloc_arraycache(int node
, int entries
,
910 int batchcount
, gfp_t gfp
)
912 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
913 struct array_cache
*nc
= NULL
;
915 nc
= kmalloc_node(memsize
, gfp
, node
);
917 * The array_cache structures contain pointers to free object.
918 * However, when such objects are allocated or transferred to another
919 * cache the pointers are not cleared and they could be counted as
920 * valid references during a kmemleak scan. Therefore, kmemleak must
921 * not scan such objects.
923 kmemleak_no_scan(nc
);
927 nc
->batchcount
= batchcount
;
929 spin_lock_init(&nc
->lock
);
934 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
936 struct page
*page
= virt_to_page(slabp
->s_mem
);
938 return PageSlabPfmemalloc(page
);
941 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
942 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
943 struct array_cache
*ac
)
945 struct kmem_list3
*l3
= cachep
->nodelists
[numa_mem_id()];
949 if (!pfmemalloc_active
)
952 spin_lock_irqsave(&l3
->list_lock
, flags
);
953 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
954 if (is_slab_pfmemalloc(slabp
))
957 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
958 if (is_slab_pfmemalloc(slabp
))
961 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
962 if (is_slab_pfmemalloc(slabp
))
965 pfmemalloc_active
= false;
967 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
970 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
971 gfp_t flags
, bool force_refill
)
974 void *objp
= ac
->entry
[--ac
->avail
];
976 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
977 if (unlikely(is_obj_pfmemalloc(objp
))) {
978 struct kmem_list3
*l3
;
980 if (gfp_pfmemalloc_allowed(flags
)) {
981 clear_obj_pfmemalloc(&objp
);
985 /* The caller cannot use PFMEMALLOC objects, find another one */
986 for (i
= 1; i
< ac
->avail
; i
++) {
987 /* If a !PFMEMALLOC object is found, swap them */
988 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
990 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
991 ac
->entry
[ac
->avail
] = objp
;
997 * If there are empty slabs on the slabs_free list and we are
998 * being forced to refill the cache, mark this one !pfmemalloc.
1000 l3
= cachep
->nodelists
[numa_mem_id()];
1001 if (!list_empty(&l3
->slabs_free
) && force_refill
) {
1002 struct slab
*slabp
= virt_to_slab(objp
);
1003 ClearPageSlabPfmemalloc(virt_to_page(slabp
->s_mem
));
1004 clear_obj_pfmemalloc(&objp
);
1005 recheck_pfmemalloc_active(cachep
, ac
);
1009 /* No !PFMEMALLOC objects available */
1017 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
1018 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
1022 if (unlikely(sk_memalloc_socks()))
1023 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
1025 objp
= ac
->entry
[--ac
->avail
];
1030 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1033 if (unlikely(pfmemalloc_active
)) {
1034 /* Some pfmemalloc slabs exist, check if this is one */
1035 struct page
*page
= virt_to_page(objp
);
1036 if (PageSlabPfmemalloc(page
))
1037 set_obj_pfmemalloc(&objp
);
1043 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1046 if (unlikely(sk_memalloc_socks()))
1047 objp
= __ac_put_obj(cachep
, ac
, objp
);
1049 ac
->entry
[ac
->avail
++] = objp
;
1053 * Transfer objects in one arraycache to another.
1054 * Locking must be handled by the caller.
1056 * Return the number of entries transferred.
1058 static int transfer_objects(struct array_cache
*to
,
1059 struct array_cache
*from
, unsigned int max
)
1061 /* Figure out how many entries to transfer */
1062 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
1067 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1068 sizeof(void *) *nr
);
1077 #define drain_alien_cache(cachep, alien) do { } while (0)
1078 #define reap_alien(cachep, l3) do { } while (0)
1080 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1082 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1085 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1089 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1094 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1100 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1101 gfp_t flags
, int nodeid
)
1106 #else /* CONFIG_NUMA */
1108 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1109 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1111 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1113 struct array_cache
**ac_ptr
;
1114 int memsize
= sizeof(void *) * nr_node_ids
;
1119 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1122 if (i
== node
|| !node_online(i
))
1124 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1126 for (i
--; i
>= 0; i
--)
1136 static void free_alien_cache(struct array_cache
**ac_ptr
)
1147 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1148 struct array_cache
*ac
, int node
)
1150 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1153 spin_lock(&rl3
->list_lock
);
1155 * Stuff objects into the remote nodes shared array first.
1156 * That way we could avoid the overhead of putting the objects
1157 * into the free lists and getting them back later.
1160 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1162 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1164 spin_unlock(&rl3
->list_lock
);
1169 * Called from cache_reap() to regularly drain alien caches round robin.
1171 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1173 int node
= __this_cpu_read(slab_reap_node
);
1176 struct array_cache
*ac
= l3
->alien
[node
];
1178 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1179 __drain_alien_cache(cachep
, ac
, node
);
1180 spin_unlock_irq(&ac
->lock
);
1185 static void drain_alien_cache(struct kmem_cache
*cachep
,
1186 struct array_cache
**alien
)
1189 struct array_cache
*ac
;
1190 unsigned long flags
;
1192 for_each_online_node(i
) {
1195 spin_lock_irqsave(&ac
->lock
, flags
);
1196 __drain_alien_cache(cachep
, ac
, i
);
1197 spin_unlock_irqrestore(&ac
->lock
, flags
);
1202 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1204 struct slab
*slabp
= virt_to_slab(objp
);
1205 int nodeid
= slabp
->nodeid
;
1206 struct kmem_list3
*l3
;
1207 struct array_cache
*alien
= NULL
;
1210 node
= numa_mem_id();
1213 * Make sure we are not freeing a object from another node to the array
1214 * cache on this cpu.
1216 if (likely(slabp
->nodeid
== node
))
1219 l3
= cachep
->nodelists
[node
];
1220 STATS_INC_NODEFREES(cachep
);
1221 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1222 alien
= l3
->alien
[nodeid
];
1223 spin_lock(&alien
->lock
);
1224 if (unlikely(alien
->avail
== alien
->limit
)) {
1225 STATS_INC_ACOVERFLOW(cachep
);
1226 __drain_alien_cache(cachep
, alien
, nodeid
);
1228 ac_put_obj(cachep
, alien
, objp
);
1229 spin_unlock(&alien
->lock
);
1231 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1232 free_block(cachep
, &objp
, 1, nodeid
);
1233 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1240 * Allocates and initializes nodelists for a node on each slab cache, used for
1241 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1242 * will be allocated off-node since memory is not yet online for the new node.
1243 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1246 * Must hold slab_mutex.
1248 static int init_cache_nodelists_node(int node
)
1250 struct kmem_cache
*cachep
;
1251 struct kmem_list3
*l3
;
1252 const int memsize
= sizeof(struct kmem_list3
);
1254 list_for_each_entry(cachep
, &slab_caches
, list
) {
1256 * Set up the size64 kmemlist for cpu before we can
1257 * begin anything. Make sure some other cpu on this
1258 * node has not already allocated this
1260 if (!cachep
->nodelists
[node
]) {
1261 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1264 kmem_list3_init(l3
);
1265 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1266 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1269 * The l3s don't come and go as CPUs come and
1270 * go. slab_mutex is sufficient
1273 cachep
->nodelists
[node
] = l3
;
1276 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1277 cachep
->nodelists
[node
]->free_limit
=
1278 (1 + nr_cpus_node(node
)) *
1279 cachep
->batchcount
+ cachep
->num
;
1280 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1285 static void __cpuinit
cpuup_canceled(long cpu
)
1287 struct kmem_cache
*cachep
;
1288 struct kmem_list3
*l3
= NULL
;
1289 int node
= cpu_to_mem(cpu
);
1290 const struct cpumask
*mask
= cpumask_of_node(node
);
1292 list_for_each_entry(cachep
, &slab_caches
, list
) {
1293 struct array_cache
*nc
;
1294 struct array_cache
*shared
;
1295 struct array_cache
**alien
;
1297 /* cpu is dead; no one can alloc from it. */
1298 nc
= cachep
->array
[cpu
];
1299 cachep
->array
[cpu
] = NULL
;
1300 l3
= cachep
->nodelists
[node
];
1303 goto free_array_cache
;
1305 spin_lock_irq(&l3
->list_lock
);
1307 /* Free limit for this kmem_list3 */
1308 l3
->free_limit
-= cachep
->batchcount
;
1310 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1312 if (!cpumask_empty(mask
)) {
1313 spin_unlock_irq(&l3
->list_lock
);
1314 goto free_array_cache
;
1317 shared
= l3
->shared
;
1319 free_block(cachep
, shared
->entry
,
1320 shared
->avail
, node
);
1327 spin_unlock_irq(&l3
->list_lock
);
1331 drain_alien_cache(cachep
, alien
);
1332 free_alien_cache(alien
);
1338 * In the previous loop, all the objects were freed to
1339 * the respective cache's slabs, now we can go ahead and
1340 * shrink each nodelist to its limit.
1342 list_for_each_entry(cachep
, &slab_caches
, list
) {
1343 l3
= cachep
->nodelists
[node
];
1346 drain_freelist(cachep
, l3
, l3
->free_objects
);
1350 static int __cpuinit
cpuup_prepare(long cpu
)
1352 struct kmem_cache
*cachep
;
1353 struct kmem_list3
*l3
= NULL
;
1354 int node
= cpu_to_mem(cpu
);
1358 * We need to do this right in the beginning since
1359 * alloc_arraycache's are going to use this list.
1360 * kmalloc_node allows us to add the slab to the right
1361 * kmem_list3 and not this cpu's kmem_list3
1363 err
= init_cache_nodelists_node(node
);
1368 * Now we can go ahead with allocating the shared arrays and
1371 list_for_each_entry(cachep
, &slab_caches
, list
) {
1372 struct array_cache
*nc
;
1373 struct array_cache
*shared
= NULL
;
1374 struct array_cache
**alien
= NULL
;
1376 nc
= alloc_arraycache(node
, cachep
->limit
,
1377 cachep
->batchcount
, GFP_KERNEL
);
1380 if (cachep
->shared
) {
1381 shared
= alloc_arraycache(node
,
1382 cachep
->shared
* cachep
->batchcount
,
1383 0xbaadf00d, GFP_KERNEL
);
1389 if (use_alien_caches
) {
1390 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1397 cachep
->array
[cpu
] = nc
;
1398 l3
= cachep
->nodelists
[node
];
1401 spin_lock_irq(&l3
->list_lock
);
1404 * We are serialised from CPU_DEAD or
1405 * CPU_UP_CANCELLED by the cpucontrol lock
1407 l3
->shared
= shared
;
1416 spin_unlock_irq(&l3
->list_lock
);
1418 free_alien_cache(alien
);
1419 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1420 slab_set_debugobj_lock_classes_node(cachep
, node
);
1422 init_node_lock_keys(node
);
1426 cpuup_canceled(cpu
);
1430 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1431 unsigned long action
, void *hcpu
)
1433 long cpu
= (long)hcpu
;
1437 case CPU_UP_PREPARE
:
1438 case CPU_UP_PREPARE_FROZEN
:
1439 mutex_lock(&slab_mutex
);
1440 err
= cpuup_prepare(cpu
);
1441 mutex_unlock(&slab_mutex
);
1444 case CPU_ONLINE_FROZEN
:
1445 start_cpu_timer(cpu
);
1447 #ifdef CONFIG_HOTPLUG_CPU
1448 case CPU_DOWN_PREPARE
:
1449 case CPU_DOWN_PREPARE_FROZEN
:
1451 * Shutdown cache reaper. Note that the slab_mutex is
1452 * held so that if cache_reap() is invoked it cannot do
1453 * anything expensive but will only modify reap_work
1454 * and reschedule the timer.
1456 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1457 /* Now the cache_reaper is guaranteed to be not running. */
1458 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1460 case CPU_DOWN_FAILED
:
1461 case CPU_DOWN_FAILED_FROZEN
:
1462 start_cpu_timer(cpu
);
1465 case CPU_DEAD_FROZEN
:
1467 * Even if all the cpus of a node are down, we don't free the
1468 * kmem_list3 of any cache. This to avoid a race between
1469 * cpu_down, and a kmalloc allocation from another cpu for
1470 * memory from the node of the cpu going down. The list3
1471 * structure is usually allocated from kmem_cache_create() and
1472 * gets destroyed at kmem_cache_destroy().
1476 case CPU_UP_CANCELED
:
1477 case CPU_UP_CANCELED_FROZEN
:
1478 mutex_lock(&slab_mutex
);
1479 cpuup_canceled(cpu
);
1480 mutex_unlock(&slab_mutex
);
1483 return notifier_from_errno(err
);
1486 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1487 &cpuup_callback
, NULL
, 0
1490 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1492 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1493 * Returns -EBUSY if all objects cannot be drained so that the node is not
1496 * Must hold slab_mutex.
1498 static int __meminit
drain_cache_nodelists_node(int node
)
1500 struct kmem_cache
*cachep
;
1503 list_for_each_entry(cachep
, &slab_caches
, list
) {
1504 struct kmem_list3
*l3
;
1506 l3
= cachep
->nodelists
[node
];
1510 drain_freelist(cachep
, l3
, l3
->free_objects
);
1512 if (!list_empty(&l3
->slabs_full
) ||
1513 !list_empty(&l3
->slabs_partial
)) {
1521 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1522 unsigned long action
, void *arg
)
1524 struct memory_notify
*mnb
= arg
;
1528 nid
= mnb
->status_change_nid
;
1533 case MEM_GOING_ONLINE
:
1534 mutex_lock(&slab_mutex
);
1535 ret
= init_cache_nodelists_node(nid
);
1536 mutex_unlock(&slab_mutex
);
1538 case MEM_GOING_OFFLINE
:
1539 mutex_lock(&slab_mutex
);
1540 ret
= drain_cache_nodelists_node(nid
);
1541 mutex_unlock(&slab_mutex
);
1545 case MEM_CANCEL_ONLINE
:
1546 case MEM_CANCEL_OFFLINE
:
1550 return notifier_from_errno(ret
);
1552 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1555 * swap the static kmem_list3 with kmalloced memory
1557 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1560 struct kmem_list3
*ptr
;
1562 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1565 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1567 * Do not assume that spinlocks can be initialized via memcpy:
1569 spin_lock_init(&ptr
->list_lock
);
1571 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1572 cachep
->nodelists
[nodeid
] = ptr
;
1576 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1577 * size of kmem_list3.
1579 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1583 for_each_online_node(node
) {
1584 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1585 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1587 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1592 * Initialisation. Called after the page allocator have been initialised and
1593 * before smp_init().
1595 void __init
kmem_cache_init(void)
1598 struct cache_sizes
*sizes
;
1599 struct cache_names
*names
;
1604 if (num_possible_nodes() == 1)
1605 use_alien_caches
= 0;
1607 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1608 kmem_list3_init(&initkmem_list3
[i
]);
1609 if (i
< MAX_NUMNODES
)
1610 cache_cache
.nodelists
[i
] = NULL
;
1612 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1615 * Fragmentation resistance on low memory - only use bigger
1616 * page orders on machines with more than 32MB of memory if
1617 * not overridden on the command line.
1619 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1620 slab_max_order
= SLAB_MAX_ORDER_HI
;
1622 /* Bootstrap is tricky, because several objects are allocated
1623 * from caches that do not exist yet:
1624 * 1) initialize the cache_cache cache: it contains the struct
1625 * kmem_cache structures of all caches, except cache_cache itself:
1626 * cache_cache is statically allocated.
1627 * Initially an __init data area is used for the head array and the
1628 * kmem_list3 structures, it's replaced with a kmalloc allocated
1629 * array at the end of the bootstrap.
1630 * 2) Create the first kmalloc cache.
1631 * The struct kmem_cache for the new cache is allocated normally.
1632 * An __init data area is used for the head array.
1633 * 3) Create the remaining kmalloc caches, with minimally sized
1635 * 4) Replace the __init data head arrays for cache_cache and the first
1636 * kmalloc cache with kmalloc allocated arrays.
1637 * 5) Replace the __init data for kmem_list3 for cache_cache and
1638 * the other cache's with kmalloc allocated memory.
1639 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1642 node
= numa_mem_id();
1644 /* 1) create the cache_cache */
1645 INIT_LIST_HEAD(&slab_caches
);
1646 list_add(&cache_cache
.list
, &slab_caches
);
1647 cache_cache
.colour_off
= cache_line_size();
1648 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1649 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1652 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1654 cache_cache
.size
= offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1655 nr_node_ids
* sizeof(struct kmem_list3
*);
1656 cache_cache
.object_size
= cache_cache
.size
;
1657 cache_cache
.size
= ALIGN(cache_cache
.size
,
1659 cache_cache
.reciprocal_buffer_size
=
1660 reciprocal_value(cache_cache
.size
);
1662 for (order
= 0; order
< MAX_ORDER
; order
++) {
1663 cache_estimate(order
, cache_cache
.size
,
1664 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1665 if (cache_cache
.num
)
1668 BUG_ON(!cache_cache
.num
);
1669 cache_cache
.gfporder
= order
;
1670 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1671 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1672 sizeof(struct slab
), cache_line_size());
1674 /* 2+3) create the kmalloc caches */
1675 sizes
= malloc_sizes
;
1676 names
= cache_names
;
1679 * Initialize the caches that provide memory for the array cache and the
1680 * kmem_list3 structures first. Without this, further allocations will
1684 sizes
[INDEX_AC
].cs_cachep
= __kmem_cache_create(names
[INDEX_AC
].name
,
1685 sizes
[INDEX_AC
].cs_size
,
1686 ARCH_KMALLOC_MINALIGN
,
1687 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1690 if (INDEX_AC
!= INDEX_L3
) {
1691 sizes
[INDEX_L3
].cs_cachep
=
1692 __kmem_cache_create(names
[INDEX_L3
].name
,
1693 sizes
[INDEX_L3
].cs_size
,
1694 ARCH_KMALLOC_MINALIGN
,
1695 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1699 slab_early_init
= 0;
1701 while (sizes
->cs_size
!= ULONG_MAX
) {
1703 * For performance, all the general caches are L1 aligned.
1704 * This should be particularly beneficial on SMP boxes, as it
1705 * eliminates "false sharing".
1706 * Note for systems short on memory removing the alignment will
1707 * allow tighter packing of the smaller caches.
1709 if (!sizes
->cs_cachep
) {
1710 sizes
->cs_cachep
= __kmem_cache_create(names
->name
,
1712 ARCH_KMALLOC_MINALIGN
,
1713 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1716 #ifdef CONFIG_ZONE_DMA
1717 sizes
->cs_dmacachep
= __kmem_cache_create(
1720 ARCH_KMALLOC_MINALIGN
,
1721 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1728 /* 4) Replace the bootstrap head arrays */
1730 struct array_cache
*ptr
;
1732 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1734 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1735 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1736 sizeof(struct arraycache_init
));
1738 * Do not assume that spinlocks can be initialized via memcpy:
1740 spin_lock_init(&ptr
->lock
);
1742 cache_cache
.array
[smp_processor_id()] = ptr
;
1744 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1746 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1747 != &initarray_generic
.cache
);
1748 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1749 sizeof(struct arraycache_init
));
1751 * Do not assume that spinlocks can be initialized via memcpy:
1753 spin_lock_init(&ptr
->lock
);
1755 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1758 /* 5) Replace the bootstrap kmem_list3's */
1762 for_each_online_node(nid
) {
1763 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1765 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1766 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1768 if (INDEX_AC
!= INDEX_L3
) {
1769 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1770 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1778 void __init
kmem_cache_init_late(void)
1780 struct kmem_cache
*cachep
;
1784 /* Annotate slab for lockdep -- annotate the malloc caches */
1787 /* 6) resize the head arrays to their final sizes */
1788 mutex_lock(&slab_mutex
);
1789 list_for_each_entry(cachep
, &slab_caches
, list
)
1790 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1792 mutex_unlock(&slab_mutex
);
1798 * Register a cpu startup notifier callback that initializes
1799 * cpu_cache_get for all new cpus
1801 register_cpu_notifier(&cpucache_notifier
);
1805 * Register a memory hotplug callback that initializes and frees
1808 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1812 * The reap timers are started later, with a module init call: That part
1813 * of the kernel is not yet operational.
1817 static int __init
cpucache_init(void)
1822 * Register the timers that return unneeded pages to the page allocator
1824 for_each_online_cpu(cpu
)
1825 start_cpu_timer(cpu
);
1831 __initcall(cpucache_init
);
1833 static noinline
void
1834 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1836 struct kmem_list3
*l3
;
1838 unsigned long flags
;
1842 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1844 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1845 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1847 for_each_online_node(node
) {
1848 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1849 unsigned long active_slabs
= 0, num_slabs
= 0;
1851 l3
= cachep
->nodelists
[node
];
1855 spin_lock_irqsave(&l3
->list_lock
, flags
);
1856 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
1857 active_objs
+= cachep
->num
;
1860 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
1861 active_objs
+= slabp
->inuse
;
1864 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
1867 free_objects
+= l3
->free_objects
;
1868 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
1870 num_slabs
+= active_slabs
;
1871 num_objs
= num_slabs
* cachep
->num
;
1873 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1874 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1880 * Interface to system's page allocator. No need to hold the cache-lock.
1882 * If we requested dmaable memory, we will get it. Even if we
1883 * did not request dmaable memory, we might get it, but that
1884 * would be relatively rare and ignorable.
1886 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1894 * Nommu uses slab's for process anonymous memory allocations, and thus
1895 * requires __GFP_COMP to properly refcount higher order allocations
1897 flags
|= __GFP_COMP
;
1900 flags
|= cachep
->allocflags
;
1901 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1902 flags
|= __GFP_RECLAIMABLE
;
1904 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1906 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1907 slab_out_of_memory(cachep
, flags
, nodeid
);
1911 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1912 if (unlikely(page
->pfmemalloc
))
1913 pfmemalloc_active
= true;
1915 nr_pages
= (1 << cachep
->gfporder
);
1916 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1917 add_zone_page_state(page_zone(page
),
1918 NR_SLAB_RECLAIMABLE
, nr_pages
);
1920 add_zone_page_state(page_zone(page
),
1921 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1922 for (i
= 0; i
< nr_pages
; i
++) {
1923 __SetPageSlab(page
+ i
);
1925 if (page
->pfmemalloc
)
1926 SetPageSlabPfmemalloc(page
+ i
);
1929 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1930 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1933 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1935 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1938 return page_address(page
);
1942 * Interface to system's page release.
1944 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1946 unsigned long i
= (1 << cachep
->gfporder
);
1947 struct page
*page
= virt_to_page(addr
);
1948 const unsigned long nr_freed
= i
;
1950 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1952 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1953 sub_zone_page_state(page_zone(page
),
1954 NR_SLAB_RECLAIMABLE
, nr_freed
);
1956 sub_zone_page_state(page_zone(page
),
1957 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1959 BUG_ON(!PageSlab(page
));
1960 __ClearPageSlabPfmemalloc(page
);
1961 __ClearPageSlab(page
);
1964 if (current
->reclaim_state
)
1965 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1966 free_pages((unsigned long)addr
, cachep
->gfporder
);
1969 static void kmem_rcu_free(struct rcu_head
*head
)
1971 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1972 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1974 kmem_freepages(cachep
, slab_rcu
->addr
);
1975 if (OFF_SLAB(cachep
))
1976 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1981 #ifdef CONFIG_DEBUG_PAGEALLOC
1982 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1983 unsigned long caller
)
1985 int size
= cachep
->object_size
;
1987 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1989 if (size
< 5 * sizeof(unsigned long))
1992 *addr
++ = 0x12345678;
1994 *addr
++ = smp_processor_id();
1995 size
-= 3 * sizeof(unsigned long);
1997 unsigned long *sptr
= &caller
;
1998 unsigned long svalue
;
2000 while (!kstack_end(sptr
)) {
2002 if (kernel_text_address(svalue
)) {
2004 size
-= sizeof(unsigned long);
2005 if (size
<= sizeof(unsigned long))
2011 *addr
++ = 0x87654321;
2015 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
2017 int size
= cachep
->object_size
;
2018 addr
= &((char *)addr
)[obj_offset(cachep
)];
2020 memset(addr
, val
, size
);
2021 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
2024 static void dump_line(char *data
, int offset
, int limit
)
2027 unsigned char error
= 0;
2030 printk(KERN_ERR
"%03x: ", offset
);
2031 for (i
= 0; i
< limit
; i
++) {
2032 if (data
[offset
+ i
] != POISON_FREE
) {
2033 error
= data
[offset
+ i
];
2037 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
2038 &data
[offset
], limit
, 1);
2040 if (bad_count
== 1) {
2041 error
^= POISON_FREE
;
2042 if (!(error
& (error
- 1))) {
2043 printk(KERN_ERR
"Single bit error detected. Probably "
2046 printk(KERN_ERR
"Run memtest86+ or a similar memory "
2049 printk(KERN_ERR
"Run a memory test tool.\n");
2058 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
2063 if (cachep
->flags
& SLAB_RED_ZONE
) {
2064 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
2065 *dbg_redzone1(cachep
, objp
),
2066 *dbg_redzone2(cachep
, objp
));
2069 if (cachep
->flags
& SLAB_STORE_USER
) {
2070 printk(KERN_ERR
"Last user: [<%p>]",
2071 *dbg_userword(cachep
, objp
));
2072 print_symbol("(%s)",
2073 (unsigned long)*dbg_userword(cachep
, objp
));
2076 realobj
= (char *)objp
+ obj_offset(cachep
);
2077 size
= cachep
->object_size
;
2078 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
2081 if (i
+ limit
> size
)
2083 dump_line(realobj
, i
, limit
);
2087 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
2093 realobj
= (char *)objp
+ obj_offset(cachep
);
2094 size
= cachep
->object_size
;
2096 for (i
= 0; i
< size
; i
++) {
2097 char exp
= POISON_FREE
;
2100 if (realobj
[i
] != exp
) {
2106 "Slab corruption (%s): %s start=%p, len=%d\n",
2107 print_tainted(), cachep
->name
, realobj
, size
);
2108 print_objinfo(cachep
, objp
, 0);
2110 /* Hexdump the affected line */
2113 if (i
+ limit
> size
)
2115 dump_line(realobj
, i
, limit
);
2118 /* Limit to 5 lines */
2124 /* Print some data about the neighboring objects, if they
2127 struct slab
*slabp
= virt_to_slab(objp
);
2130 objnr
= obj_to_index(cachep
, slabp
, objp
);
2132 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
2133 realobj
= (char *)objp
+ obj_offset(cachep
);
2134 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
2136 print_objinfo(cachep
, objp
, 2);
2138 if (objnr
+ 1 < cachep
->num
) {
2139 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
2140 realobj
= (char *)objp
+ obj_offset(cachep
);
2141 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
2143 print_objinfo(cachep
, objp
, 2);
2150 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2153 for (i
= 0; i
< cachep
->num
; i
++) {
2154 void *objp
= index_to_obj(cachep
, slabp
, i
);
2156 if (cachep
->flags
& SLAB_POISON
) {
2157 #ifdef CONFIG_DEBUG_PAGEALLOC
2158 if (cachep
->size
% PAGE_SIZE
== 0 &&
2160 kernel_map_pages(virt_to_page(objp
),
2161 cachep
->size
/ PAGE_SIZE
, 1);
2163 check_poison_obj(cachep
, objp
);
2165 check_poison_obj(cachep
, objp
);
2168 if (cachep
->flags
& SLAB_RED_ZONE
) {
2169 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2170 slab_error(cachep
, "start of a freed object "
2172 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2173 slab_error(cachep
, "end of a freed object "
2179 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2185 * slab_destroy - destroy and release all objects in a slab
2186 * @cachep: cache pointer being destroyed
2187 * @slabp: slab pointer being destroyed
2189 * Destroy all the objs in a slab, and release the mem back to the system.
2190 * Before calling the slab must have been unlinked from the cache. The
2191 * cache-lock is not held/needed.
2193 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2195 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2197 slab_destroy_debugcheck(cachep
, slabp
);
2198 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2199 struct slab_rcu
*slab_rcu
;
2201 slab_rcu
= (struct slab_rcu
*)slabp
;
2202 slab_rcu
->cachep
= cachep
;
2203 slab_rcu
->addr
= addr
;
2204 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2206 kmem_freepages(cachep
, addr
);
2207 if (OFF_SLAB(cachep
))
2208 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2212 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
2215 struct kmem_list3
*l3
;
2217 for_each_online_cpu(i
)
2218 kfree(cachep
->array
[i
]);
2220 /* NUMA: free the list3 structures */
2221 for_each_online_node(i
) {
2222 l3
= cachep
->nodelists
[i
];
2225 free_alien_cache(l3
->alien
);
2229 kmem_cache_free(&cache_cache
, cachep
);
2234 * calculate_slab_order - calculate size (page order) of slabs
2235 * @cachep: pointer to the cache that is being created
2236 * @size: size of objects to be created in this cache.
2237 * @align: required alignment for the objects.
2238 * @flags: slab allocation flags
2240 * Also calculates the number of objects per slab.
2242 * This could be made much more intelligent. For now, try to avoid using
2243 * high order pages for slabs. When the gfp() functions are more friendly
2244 * towards high-order requests, this should be changed.
2246 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2247 size_t size
, size_t align
, unsigned long flags
)
2249 unsigned long offslab_limit
;
2250 size_t left_over
= 0;
2253 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2257 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2261 if (flags
& CFLGS_OFF_SLAB
) {
2263 * Max number of objs-per-slab for caches which
2264 * use off-slab slabs. Needed to avoid a possible
2265 * looping condition in cache_grow().
2267 offslab_limit
= size
- sizeof(struct slab
);
2268 offslab_limit
/= sizeof(kmem_bufctl_t
);
2270 if (num
> offslab_limit
)
2274 /* Found something acceptable - save it away */
2276 cachep
->gfporder
= gfporder
;
2277 left_over
= remainder
;
2280 * A VFS-reclaimable slab tends to have most allocations
2281 * as GFP_NOFS and we really don't want to have to be allocating
2282 * higher-order pages when we are unable to shrink dcache.
2284 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2288 * Large number of objects is good, but very large slabs are
2289 * currently bad for the gfp()s.
2291 if (gfporder
>= slab_max_order
)
2295 * Acceptable internal fragmentation?
2297 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2303 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2305 if (slab_state
>= FULL
)
2306 return enable_cpucache(cachep
, gfp
);
2308 if (slab_state
== DOWN
) {
2310 * Note: the first kmem_cache_create must create the cache
2311 * that's used by kmalloc(24), otherwise the creation of
2312 * further caches will BUG().
2314 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2317 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2318 * the first cache, then we need to set up all its list3s,
2319 * otherwise the creation of further caches will BUG().
2321 set_up_list3s(cachep
, SIZE_AC
);
2322 if (INDEX_AC
== INDEX_L3
)
2323 slab_state
= PARTIAL_L3
;
2325 slab_state
= PARTIAL_ARRAYCACHE
;
2327 cachep
->array
[smp_processor_id()] =
2328 kmalloc(sizeof(struct arraycache_init
), gfp
);
2330 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2331 set_up_list3s(cachep
, SIZE_L3
);
2332 slab_state
= PARTIAL_L3
;
2335 for_each_online_node(node
) {
2336 cachep
->nodelists
[node
] =
2337 kmalloc_node(sizeof(struct kmem_list3
),
2339 BUG_ON(!cachep
->nodelists
[node
]);
2340 kmem_list3_init(cachep
->nodelists
[node
]);
2344 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2345 jiffies
+ REAPTIMEOUT_LIST3
+
2346 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2348 cpu_cache_get(cachep
)->avail
= 0;
2349 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2350 cpu_cache_get(cachep
)->batchcount
= 1;
2351 cpu_cache_get(cachep
)->touched
= 0;
2352 cachep
->batchcount
= 1;
2353 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2358 * __kmem_cache_create - Create a cache.
2359 * @name: A string which is used in /proc/slabinfo to identify this cache.
2360 * @size: The size of objects to be created in this cache.
2361 * @align: The required alignment for the objects.
2362 * @flags: SLAB flags
2363 * @ctor: A constructor for the objects.
2365 * Returns a ptr to the cache on success, NULL on failure.
2366 * Cannot be called within a int, but can be interrupted.
2367 * The @ctor is run when new pages are allocated by the cache.
2369 * @name must be valid until the cache is destroyed. This implies that
2370 * the module calling this has to destroy the cache before getting unloaded.
2374 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2375 * to catch references to uninitialised memory.
2377 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2378 * for buffer overruns.
2380 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2381 * cacheline. This can be beneficial if you're counting cycles as closely
2385 __kmem_cache_create (const char *name
, size_t size
, size_t align
,
2386 unsigned long flags
, void (*ctor
)(void *))
2388 size_t left_over
, slab_size
, ralign
;
2389 struct kmem_cache
*cachep
= NULL
;
2395 * Enable redzoning and last user accounting, except for caches with
2396 * large objects, if the increased size would increase the object size
2397 * above the next power of two: caches with object sizes just above a
2398 * power of two have a significant amount of internal fragmentation.
2400 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2401 2 * sizeof(unsigned long long)))
2402 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2403 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2404 flags
|= SLAB_POISON
;
2406 if (flags
& SLAB_DESTROY_BY_RCU
)
2407 BUG_ON(flags
& SLAB_POISON
);
2410 * Always checks flags, a caller might be expecting debug support which
2413 BUG_ON(flags
& ~CREATE_MASK
);
2416 * Check that size is in terms of words. This is needed to avoid
2417 * unaligned accesses for some archs when redzoning is used, and makes
2418 * sure any on-slab bufctl's are also correctly aligned.
2420 if (size
& (BYTES_PER_WORD
- 1)) {
2421 size
+= (BYTES_PER_WORD
- 1);
2422 size
&= ~(BYTES_PER_WORD
- 1);
2425 /* calculate the final buffer alignment: */
2427 /* 1) arch recommendation: can be overridden for debug */
2428 if (flags
& SLAB_HWCACHE_ALIGN
) {
2430 * Default alignment: as specified by the arch code. Except if
2431 * an object is really small, then squeeze multiple objects into
2434 ralign
= cache_line_size();
2435 while (size
<= ralign
/ 2)
2438 ralign
= BYTES_PER_WORD
;
2442 * Redzoning and user store require word alignment or possibly larger.
2443 * Note this will be overridden by architecture or caller mandated
2444 * alignment if either is greater than BYTES_PER_WORD.
2446 if (flags
& SLAB_STORE_USER
)
2447 ralign
= BYTES_PER_WORD
;
2449 if (flags
& SLAB_RED_ZONE
) {
2450 ralign
= REDZONE_ALIGN
;
2451 /* If redzoning, ensure that the second redzone is suitably
2452 * aligned, by adjusting the object size accordingly. */
2453 size
+= REDZONE_ALIGN
- 1;
2454 size
&= ~(REDZONE_ALIGN
- 1);
2457 /* 2) arch mandated alignment */
2458 if (ralign
< ARCH_SLAB_MINALIGN
) {
2459 ralign
= ARCH_SLAB_MINALIGN
;
2461 /* 3) caller mandated alignment */
2462 if (ralign
< align
) {
2465 /* disable debug if necessary */
2466 if (ralign
> __alignof__(unsigned long long))
2467 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2473 if (slab_is_available())
2478 /* Get cache's description obj. */
2479 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2483 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
2484 cachep
->object_size
= size
;
2485 cachep
->align
= align
;
2489 * Both debugging options require word-alignment which is calculated
2492 if (flags
& SLAB_RED_ZONE
) {
2493 /* add space for red zone words */
2494 cachep
->obj_offset
+= sizeof(unsigned long long);
2495 size
+= 2 * sizeof(unsigned long long);
2497 if (flags
& SLAB_STORE_USER
) {
2498 /* user store requires one word storage behind the end of
2499 * the real object. But if the second red zone needs to be
2500 * aligned to 64 bits, we must allow that much space.
2502 if (flags
& SLAB_RED_ZONE
)
2503 size
+= REDZONE_ALIGN
;
2505 size
+= BYTES_PER_WORD
;
2507 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2508 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2509 && cachep
->object_size
> cache_line_size() && ALIGN(size
, align
) < PAGE_SIZE
) {
2510 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, align
);
2517 * Determine if the slab management is 'on' or 'off' slab.
2518 * (bootstrapping cannot cope with offslab caches so don't do
2519 * it too early on. Always use on-slab management when
2520 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2522 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2523 !(flags
& SLAB_NOLEAKTRACE
))
2525 * Size is large, assume best to place the slab management obj
2526 * off-slab (should allow better packing of objs).
2528 flags
|= CFLGS_OFF_SLAB
;
2530 size
= ALIGN(size
, align
);
2532 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2536 "kmem_cache_create: couldn't create cache %s.\n", name
);
2537 kmem_cache_free(&cache_cache
, cachep
);
2540 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2541 + sizeof(struct slab
), align
);
2544 * If the slab has been placed off-slab, and we have enough space then
2545 * move it on-slab. This is at the expense of any extra colouring.
2547 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2548 flags
&= ~CFLGS_OFF_SLAB
;
2549 left_over
-= slab_size
;
2552 if (flags
& CFLGS_OFF_SLAB
) {
2553 /* really off slab. No need for manual alignment */
2555 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2557 #ifdef CONFIG_PAGE_POISONING
2558 /* If we're going to use the generic kernel_map_pages()
2559 * poisoning, then it's going to smash the contents of
2560 * the redzone and userword anyhow, so switch them off.
2562 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2563 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2567 cachep
->colour_off
= cache_line_size();
2568 /* Offset must be a multiple of the alignment. */
2569 if (cachep
->colour_off
< align
)
2570 cachep
->colour_off
= align
;
2571 cachep
->colour
= left_over
/ cachep
->colour_off
;
2572 cachep
->slab_size
= slab_size
;
2573 cachep
->flags
= flags
;
2574 cachep
->allocflags
= 0;
2575 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2576 cachep
->allocflags
|= GFP_DMA
;
2577 cachep
->size
= size
;
2578 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2580 if (flags
& CFLGS_OFF_SLAB
) {
2581 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2583 * This is a possibility for one of the malloc_sizes caches.
2584 * But since we go off slab only for object size greater than
2585 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2586 * this should not happen at all.
2587 * But leave a BUG_ON for some lucky dude.
2589 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2591 cachep
->ctor
= ctor
;
2592 cachep
->name
= name
;
2594 if (setup_cpu_cache(cachep
, gfp
)) {
2595 __kmem_cache_destroy(cachep
);
2599 if (flags
& SLAB_DEBUG_OBJECTS
) {
2601 * Would deadlock through slab_destroy()->call_rcu()->
2602 * debug_object_activate()->kmem_cache_alloc().
2604 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2606 slab_set_debugobj_lock_classes(cachep
);
2609 /* cache setup completed, link it into the list */
2610 list_add(&cachep
->list
, &slab_caches
);
2615 static void check_irq_off(void)
2617 BUG_ON(!irqs_disabled());
2620 static void check_irq_on(void)
2622 BUG_ON(irqs_disabled());
2625 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2629 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2633 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2637 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2642 #define check_irq_off() do { } while(0)
2643 #define check_irq_on() do { } while(0)
2644 #define check_spinlock_acquired(x) do { } while(0)
2645 #define check_spinlock_acquired_node(x, y) do { } while(0)
2648 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2649 struct array_cache
*ac
,
2650 int force
, int node
);
2652 static void do_drain(void *arg
)
2654 struct kmem_cache
*cachep
= arg
;
2655 struct array_cache
*ac
;
2656 int node
= numa_mem_id();
2659 ac
= cpu_cache_get(cachep
);
2660 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2661 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2662 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2666 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2668 struct kmem_list3
*l3
;
2671 on_each_cpu(do_drain
, cachep
, 1);
2673 for_each_online_node(node
) {
2674 l3
= cachep
->nodelists
[node
];
2675 if (l3
&& l3
->alien
)
2676 drain_alien_cache(cachep
, l3
->alien
);
2679 for_each_online_node(node
) {
2680 l3
= cachep
->nodelists
[node
];
2682 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2687 * Remove slabs from the list of free slabs.
2688 * Specify the number of slabs to drain in tofree.
2690 * Returns the actual number of slabs released.
2692 static int drain_freelist(struct kmem_cache
*cache
,
2693 struct kmem_list3
*l3
, int tofree
)
2695 struct list_head
*p
;
2700 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2702 spin_lock_irq(&l3
->list_lock
);
2703 p
= l3
->slabs_free
.prev
;
2704 if (p
== &l3
->slabs_free
) {
2705 spin_unlock_irq(&l3
->list_lock
);
2709 slabp
= list_entry(p
, struct slab
, list
);
2711 BUG_ON(slabp
->inuse
);
2713 list_del(&slabp
->list
);
2715 * Safe to drop the lock. The slab is no longer linked
2718 l3
->free_objects
-= cache
->num
;
2719 spin_unlock_irq(&l3
->list_lock
);
2720 slab_destroy(cache
, slabp
);
2727 /* Called with slab_mutex held to protect against cpu hotplug */
2728 static int __cache_shrink(struct kmem_cache
*cachep
)
2731 struct kmem_list3
*l3
;
2733 drain_cpu_caches(cachep
);
2736 for_each_online_node(i
) {
2737 l3
= cachep
->nodelists
[i
];
2741 drain_freelist(cachep
, l3
, l3
->free_objects
);
2743 ret
+= !list_empty(&l3
->slabs_full
) ||
2744 !list_empty(&l3
->slabs_partial
);
2746 return (ret
? 1 : 0);
2750 * kmem_cache_shrink - Shrink a cache.
2751 * @cachep: The cache to shrink.
2753 * Releases as many slabs as possible for a cache.
2754 * To help debugging, a zero exit status indicates all slabs were released.
2756 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2759 BUG_ON(!cachep
|| in_interrupt());
2762 mutex_lock(&slab_mutex
);
2763 ret
= __cache_shrink(cachep
);
2764 mutex_unlock(&slab_mutex
);
2768 EXPORT_SYMBOL(kmem_cache_shrink
);
2771 * kmem_cache_destroy - delete a cache
2772 * @cachep: the cache to destroy
2774 * Remove a &struct kmem_cache object from the slab cache.
2776 * It is expected this function will be called by a module when it is
2777 * unloaded. This will remove the cache completely, and avoid a duplicate
2778 * cache being allocated each time a module is loaded and unloaded, if the
2779 * module doesn't have persistent in-kernel storage across loads and unloads.
2781 * The cache must be empty before calling this function.
2783 * The caller must guarantee that no one will allocate memory from the cache
2784 * during the kmem_cache_destroy().
2786 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2788 BUG_ON(!cachep
|| in_interrupt());
2790 /* Find the cache in the chain of caches. */
2792 mutex_lock(&slab_mutex
);
2794 * the chain is never empty, cache_cache is never destroyed
2796 list_del(&cachep
->list
);
2797 if (__cache_shrink(cachep
)) {
2798 slab_error(cachep
, "Can't free all objects");
2799 list_add(&cachep
->list
, &slab_caches
);
2800 mutex_unlock(&slab_mutex
);
2805 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2808 __kmem_cache_destroy(cachep
);
2809 mutex_unlock(&slab_mutex
);
2812 EXPORT_SYMBOL(kmem_cache_destroy
);
2815 * Get the memory for a slab management obj.
2816 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2817 * always come from malloc_sizes caches. The slab descriptor cannot
2818 * come from the same cache which is getting created because,
2819 * when we are searching for an appropriate cache for these
2820 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2821 * If we are creating a malloc_sizes cache here it would not be visible to
2822 * kmem_find_general_cachep till the initialization is complete.
2823 * Hence we cannot have slabp_cache same as the original cache.
2825 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2826 int colour_off
, gfp_t local_flags
,
2831 if (OFF_SLAB(cachep
)) {
2832 /* Slab management obj is off-slab. */
2833 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2834 local_flags
, nodeid
);
2836 * If the first object in the slab is leaked (it's allocated
2837 * but no one has a reference to it), we want to make sure
2838 * kmemleak does not treat the ->s_mem pointer as a reference
2839 * to the object. Otherwise we will not report the leak.
2841 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2846 slabp
= objp
+ colour_off
;
2847 colour_off
+= cachep
->slab_size
;
2850 slabp
->colouroff
= colour_off
;
2851 slabp
->s_mem
= objp
+ colour_off
;
2852 slabp
->nodeid
= nodeid
;
2857 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2859 return (kmem_bufctl_t
*) (slabp
+ 1);
2862 static void cache_init_objs(struct kmem_cache
*cachep
,
2867 for (i
= 0; i
< cachep
->num
; i
++) {
2868 void *objp
= index_to_obj(cachep
, slabp
, i
);
2870 /* need to poison the objs? */
2871 if (cachep
->flags
& SLAB_POISON
)
2872 poison_obj(cachep
, objp
, POISON_FREE
);
2873 if (cachep
->flags
& SLAB_STORE_USER
)
2874 *dbg_userword(cachep
, objp
) = NULL
;
2876 if (cachep
->flags
& SLAB_RED_ZONE
) {
2877 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2878 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2881 * Constructors are not allowed to allocate memory from the same
2882 * cache which they are a constructor for. Otherwise, deadlock.
2883 * They must also be threaded.
2885 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2886 cachep
->ctor(objp
+ obj_offset(cachep
));
2888 if (cachep
->flags
& SLAB_RED_ZONE
) {
2889 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2890 slab_error(cachep
, "constructor overwrote the"
2891 " end of an object");
2892 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2893 slab_error(cachep
, "constructor overwrote the"
2894 " start of an object");
2896 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2897 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2898 kernel_map_pages(virt_to_page(objp
),
2899 cachep
->size
/ PAGE_SIZE
, 0);
2904 slab_bufctl(slabp
)[i
] = i
+ 1;
2906 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2909 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2911 if (CONFIG_ZONE_DMA_FLAG
) {
2912 if (flags
& GFP_DMA
)
2913 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2915 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2919 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2922 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2926 next
= slab_bufctl(slabp
)[slabp
->free
];
2928 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2929 WARN_ON(slabp
->nodeid
!= nodeid
);
2936 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2937 void *objp
, int nodeid
)
2939 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2942 /* Verify that the slab belongs to the intended node */
2943 WARN_ON(slabp
->nodeid
!= nodeid
);
2945 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2946 printk(KERN_ERR
"slab: double free detected in cache "
2947 "'%s', objp %p\n", cachep
->name
, objp
);
2951 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2952 slabp
->free
= objnr
;
2957 * Map pages beginning at addr to the given cache and slab. This is required
2958 * for the slab allocator to be able to lookup the cache and slab of a
2959 * virtual address for kfree, ksize, and slab debugging.
2961 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2967 page
= virt_to_page(addr
);
2970 if (likely(!PageCompound(page
)))
2971 nr_pages
<<= cache
->gfporder
;
2974 page
->slab_cache
= cache
;
2975 page
->slab_page
= slab
;
2977 } while (--nr_pages
);
2981 * Grow (by 1) the number of slabs within a cache. This is called by
2982 * kmem_cache_alloc() when there are no active objs left in a cache.
2984 static int cache_grow(struct kmem_cache
*cachep
,
2985 gfp_t flags
, int nodeid
, void *objp
)
2990 struct kmem_list3
*l3
;
2993 * Be lazy and only check for valid flags here, keeping it out of the
2994 * critical path in kmem_cache_alloc().
2996 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2997 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2999 /* Take the l3 list lock to change the colour_next on this node */
3001 l3
= cachep
->nodelists
[nodeid
];
3002 spin_lock(&l3
->list_lock
);
3004 /* Get colour for the slab, and cal the next value. */
3005 offset
= l3
->colour_next
;
3007 if (l3
->colour_next
>= cachep
->colour
)
3008 l3
->colour_next
= 0;
3009 spin_unlock(&l3
->list_lock
);
3011 offset
*= cachep
->colour_off
;
3013 if (local_flags
& __GFP_WAIT
)
3017 * The test for missing atomic flag is performed here, rather than
3018 * the more obvious place, simply to reduce the critical path length
3019 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
3020 * will eventually be caught here (where it matters).
3022 kmem_flagcheck(cachep
, flags
);
3025 * Get mem for the objs. Attempt to allocate a physical page from
3029 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
3033 /* Get slab management. */
3034 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
3035 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
3039 slab_map_pages(cachep
, slabp
, objp
);
3041 cache_init_objs(cachep
, slabp
);
3043 if (local_flags
& __GFP_WAIT
)
3044 local_irq_disable();
3046 spin_lock(&l3
->list_lock
);
3048 /* Make slab active. */
3049 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
3050 STATS_INC_GROWN(cachep
);
3051 l3
->free_objects
+= cachep
->num
;
3052 spin_unlock(&l3
->list_lock
);
3055 kmem_freepages(cachep
, objp
);
3057 if (local_flags
& __GFP_WAIT
)
3058 local_irq_disable();
3065 * Perform extra freeing checks:
3066 * - detect bad pointers.
3067 * - POISON/RED_ZONE checking
3069 static void kfree_debugcheck(const void *objp
)
3071 if (!virt_addr_valid(objp
)) {
3072 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
3073 (unsigned long)objp
);
3078 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
3080 unsigned long long redzone1
, redzone2
;
3082 redzone1
= *dbg_redzone1(cache
, obj
);
3083 redzone2
= *dbg_redzone2(cache
, obj
);
3088 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
3091 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
3092 slab_error(cache
, "double free detected");
3094 slab_error(cache
, "memory outside object was overwritten");
3096 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3097 obj
, redzone1
, redzone2
);
3100 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
3107 BUG_ON(virt_to_cache(objp
) != cachep
);
3109 objp
-= obj_offset(cachep
);
3110 kfree_debugcheck(objp
);
3111 page
= virt_to_head_page(objp
);
3113 slabp
= page
->slab_page
;
3115 if (cachep
->flags
& SLAB_RED_ZONE
) {
3116 verify_redzone_free(cachep
, objp
);
3117 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
3118 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
3120 if (cachep
->flags
& SLAB_STORE_USER
)
3121 *dbg_userword(cachep
, objp
) = caller
;
3123 objnr
= obj_to_index(cachep
, slabp
, objp
);
3125 BUG_ON(objnr
>= cachep
->num
);
3126 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
3128 #ifdef CONFIG_DEBUG_SLAB_LEAK
3129 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
3131 if (cachep
->flags
& SLAB_POISON
) {
3132 #ifdef CONFIG_DEBUG_PAGEALLOC
3133 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
3134 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
3135 kernel_map_pages(virt_to_page(objp
),
3136 cachep
->size
/ PAGE_SIZE
, 0);
3138 poison_obj(cachep
, objp
, POISON_FREE
);
3141 poison_obj(cachep
, objp
, POISON_FREE
);
3147 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
3152 /* Check slab's freelist to see if this obj is there. */
3153 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
3155 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3158 if (entries
!= cachep
->num
- slabp
->inuse
) {
3160 printk(KERN_ERR
"slab: Internal list corruption detected in "
3161 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3162 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
3164 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3165 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3171 #define kfree_debugcheck(x) do { } while(0)
3172 #define cache_free_debugcheck(x,objp,z) (objp)
3173 #define check_slabp(x,y) do { } while(0)
3176 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
3180 struct kmem_list3
*l3
;
3181 struct array_cache
*ac
;
3185 node
= numa_mem_id();
3186 if (unlikely(force_refill
))
3189 ac
= cpu_cache_get(cachep
);
3190 batchcount
= ac
->batchcount
;
3191 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3193 * If there was little recent activity on this cache, then
3194 * perform only a partial refill. Otherwise we could generate
3197 batchcount
= BATCHREFILL_LIMIT
;
3199 l3
= cachep
->nodelists
[node
];
3201 BUG_ON(ac
->avail
> 0 || !l3
);
3202 spin_lock(&l3
->list_lock
);
3204 /* See if we can refill from the shared array */
3205 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3206 l3
->shared
->touched
= 1;
3210 while (batchcount
> 0) {
3211 struct list_head
*entry
;
3213 /* Get slab alloc is to come from. */
3214 entry
= l3
->slabs_partial
.next
;
3215 if (entry
== &l3
->slabs_partial
) {
3216 l3
->free_touched
= 1;
3217 entry
= l3
->slabs_free
.next
;
3218 if (entry
== &l3
->slabs_free
)
3222 slabp
= list_entry(entry
, struct slab
, list
);
3223 check_slabp(cachep
, slabp
);
3224 check_spinlock_acquired(cachep
);
3227 * The slab was either on partial or free list so
3228 * there must be at least one object available for
3231 BUG_ON(slabp
->inuse
>= cachep
->num
);
3233 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3234 STATS_INC_ALLOCED(cachep
);
3235 STATS_INC_ACTIVE(cachep
);
3236 STATS_SET_HIGH(cachep
);
3238 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
3241 check_slabp(cachep
, slabp
);
3243 /* move slabp to correct slabp list: */
3244 list_del(&slabp
->list
);
3245 if (slabp
->free
== BUFCTL_END
)
3246 list_add(&slabp
->list
, &l3
->slabs_full
);
3248 list_add(&slabp
->list
, &l3
->slabs_partial
);
3252 l3
->free_objects
-= ac
->avail
;
3254 spin_unlock(&l3
->list_lock
);
3256 if (unlikely(!ac
->avail
)) {
3259 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3261 /* cache_grow can reenable interrupts, then ac could change. */
3262 ac
= cpu_cache_get(cachep
);
3264 /* no objects in sight? abort */
3265 if (!x
&& (ac
->avail
== 0 || force_refill
))
3268 if (!ac
->avail
) /* objects refilled by interrupt? */
3273 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
3276 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3279 might_sleep_if(flags
& __GFP_WAIT
);
3281 kmem_flagcheck(cachep
, flags
);
3286 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3287 gfp_t flags
, void *objp
, void *caller
)
3291 if (cachep
->flags
& SLAB_POISON
) {
3292 #ifdef CONFIG_DEBUG_PAGEALLOC
3293 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3294 kernel_map_pages(virt_to_page(objp
),
3295 cachep
->size
/ PAGE_SIZE
, 1);
3297 check_poison_obj(cachep
, objp
);
3299 check_poison_obj(cachep
, objp
);
3301 poison_obj(cachep
, objp
, POISON_INUSE
);
3303 if (cachep
->flags
& SLAB_STORE_USER
)
3304 *dbg_userword(cachep
, objp
) = caller
;
3306 if (cachep
->flags
& SLAB_RED_ZONE
) {
3307 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3308 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3309 slab_error(cachep
, "double free, or memory outside"
3310 " object was overwritten");
3312 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3313 objp
, *dbg_redzone1(cachep
, objp
),
3314 *dbg_redzone2(cachep
, objp
));
3316 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3317 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3319 #ifdef CONFIG_DEBUG_SLAB_LEAK
3324 slabp
= virt_to_head_page(objp
)->slab_page
;
3325 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3326 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3329 objp
+= obj_offset(cachep
);
3330 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3332 if (ARCH_SLAB_MINALIGN
&&
3333 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3334 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3335 objp
, (int)ARCH_SLAB_MINALIGN
);
3340 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3343 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3345 if (cachep
== &cache_cache
)
3348 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3351 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3354 struct array_cache
*ac
;
3355 bool force_refill
= false;
3359 ac
= cpu_cache_get(cachep
);
3360 if (likely(ac
->avail
)) {
3362 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3365 * Allow for the possibility all avail objects are not allowed
3366 * by the current flags
3369 STATS_INC_ALLOCHIT(cachep
);
3372 force_refill
= true;
3375 STATS_INC_ALLOCMISS(cachep
);
3376 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3378 * the 'ac' may be updated by cache_alloc_refill(),
3379 * and kmemleak_erase() requires its correct value.
3381 ac
= cpu_cache_get(cachep
);
3385 * To avoid a false negative, if an object that is in one of the
3386 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3387 * treat the array pointers as a reference to the object.
3390 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3396 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3398 * If we are in_interrupt, then process context, including cpusets and
3399 * mempolicy, may not apply and should not be used for allocation policy.
3401 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3403 int nid_alloc
, nid_here
;
3405 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3407 nid_alloc
= nid_here
= numa_mem_id();
3408 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3409 nid_alloc
= cpuset_slab_spread_node();
3410 else if (current
->mempolicy
)
3411 nid_alloc
= slab_node();
3412 if (nid_alloc
!= nid_here
)
3413 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3418 * Fallback function if there was no memory available and no objects on a
3419 * certain node and fall back is permitted. First we scan all the
3420 * available nodelists for available objects. If that fails then we
3421 * perform an allocation without specifying a node. This allows the page
3422 * allocator to do its reclaim / fallback magic. We then insert the
3423 * slab into the proper nodelist and then allocate from it.
3425 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3427 struct zonelist
*zonelist
;
3431 enum zone_type high_zoneidx
= gfp_zone(flags
);
3434 unsigned int cpuset_mems_cookie
;
3436 if (flags
& __GFP_THISNODE
)
3439 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3442 cpuset_mems_cookie
= get_mems_allowed();
3443 zonelist
= node_zonelist(slab_node(), flags
);
3447 * Look through allowed nodes for objects available
3448 * from existing per node queues.
3450 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3451 nid
= zone_to_nid(zone
);
3453 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3454 cache
->nodelists
[nid
] &&
3455 cache
->nodelists
[nid
]->free_objects
) {
3456 obj
= ____cache_alloc_node(cache
,
3457 flags
| GFP_THISNODE
, nid
);
3465 * This allocation will be performed within the constraints
3466 * of the current cpuset / memory policy requirements.
3467 * We may trigger various forms of reclaim on the allowed
3468 * set and go into memory reserves if necessary.
3470 if (local_flags
& __GFP_WAIT
)
3472 kmem_flagcheck(cache
, flags
);
3473 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3474 if (local_flags
& __GFP_WAIT
)
3475 local_irq_disable();
3478 * Insert into the appropriate per node queues
3480 nid
= page_to_nid(virt_to_page(obj
));
3481 if (cache_grow(cache
, flags
, nid
, obj
)) {
3482 obj
= ____cache_alloc_node(cache
,
3483 flags
| GFP_THISNODE
, nid
);
3486 * Another processor may allocate the
3487 * objects in the slab since we are
3488 * not holding any locks.
3492 /* cache_grow already freed obj */
3498 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3504 * A interface to enable slab creation on nodeid
3506 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3509 struct list_head
*entry
;
3511 struct kmem_list3
*l3
;
3515 l3
= cachep
->nodelists
[nodeid
];
3520 spin_lock(&l3
->list_lock
);
3521 entry
= l3
->slabs_partial
.next
;
3522 if (entry
== &l3
->slabs_partial
) {
3523 l3
->free_touched
= 1;
3524 entry
= l3
->slabs_free
.next
;
3525 if (entry
== &l3
->slabs_free
)
3529 slabp
= list_entry(entry
, struct slab
, list
);
3530 check_spinlock_acquired_node(cachep
, nodeid
);
3531 check_slabp(cachep
, slabp
);
3533 STATS_INC_NODEALLOCS(cachep
);
3534 STATS_INC_ACTIVE(cachep
);
3535 STATS_SET_HIGH(cachep
);
3537 BUG_ON(slabp
->inuse
== cachep
->num
);
3539 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3540 check_slabp(cachep
, slabp
);
3542 /* move slabp to correct slabp list: */
3543 list_del(&slabp
->list
);
3545 if (slabp
->free
== BUFCTL_END
)
3546 list_add(&slabp
->list
, &l3
->slabs_full
);
3548 list_add(&slabp
->list
, &l3
->slabs_partial
);
3550 spin_unlock(&l3
->list_lock
);
3554 spin_unlock(&l3
->list_lock
);
3555 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3559 return fallback_alloc(cachep
, flags
);
3566 * kmem_cache_alloc_node - Allocate an object on the specified node
3567 * @cachep: The cache to allocate from.
3568 * @flags: See kmalloc().
3569 * @nodeid: node number of the target node.
3570 * @caller: return address of caller, used for debug information
3572 * Identical to kmem_cache_alloc but it will allocate memory on the given
3573 * node, which can improve the performance for cpu bound structures.
3575 * Fallback to other node is possible if __GFP_THISNODE is not set.
3577 static __always_inline
void *
3578 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3581 unsigned long save_flags
;
3583 int slab_node
= numa_mem_id();
3585 flags
&= gfp_allowed_mask
;
3587 lockdep_trace_alloc(flags
);
3589 if (slab_should_failslab(cachep
, flags
))
3592 cache_alloc_debugcheck_before(cachep
, flags
);
3593 local_irq_save(save_flags
);
3595 if (nodeid
== NUMA_NO_NODE
)
3598 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3599 /* Node not bootstrapped yet */
3600 ptr
= fallback_alloc(cachep
, flags
);
3604 if (nodeid
== slab_node
) {
3606 * Use the locally cached objects if possible.
3607 * However ____cache_alloc does not allow fallback
3608 * to other nodes. It may fail while we still have
3609 * objects on other nodes available.
3611 ptr
= ____cache_alloc(cachep
, flags
);
3615 /* ___cache_alloc_node can fall back to other nodes */
3616 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3618 local_irq_restore(save_flags
);
3619 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3620 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3624 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3626 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3627 memset(ptr
, 0, cachep
->object_size
);
3632 static __always_inline
void *
3633 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3637 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3638 objp
= alternate_node_alloc(cache
, flags
);
3642 objp
= ____cache_alloc(cache
, flags
);
3645 * We may just have run out of memory on the local node.
3646 * ____cache_alloc_node() knows how to locate memory on other nodes
3649 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3656 static __always_inline
void *
3657 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3659 return ____cache_alloc(cachep
, flags
);
3662 #endif /* CONFIG_NUMA */
3664 static __always_inline
void *
3665 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3667 unsigned long save_flags
;
3670 flags
&= gfp_allowed_mask
;
3672 lockdep_trace_alloc(flags
);
3674 if (slab_should_failslab(cachep
, flags
))
3677 cache_alloc_debugcheck_before(cachep
, flags
);
3678 local_irq_save(save_flags
);
3679 objp
= __do_cache_alloc(cachep
, flags
);
3680 local_irq_restore(save_flags
);
3681 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3682 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3687 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3689 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3690 memset(objp
, 0, cachep
->object_size
);
3696 * Caller needs to acquire correct kmem_list's list_lock
3698 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3702 struct kmem_list3
*l3
;
3704 for (i
= 0; i
< nr_objects
; i
++) {
3708 clear_obj_pfmemalloc(&objpp
[i
]);
3711 slabp
= virt_to_slab(objp
);
3712 l3
= cachep
->nodelists
[node
];
3713 list_del(&slabp
->list
);
3714 check_spinlock_acquired_node(cachep
, node
);
3715 check_slabp(cachep
, slabp
);
3716 slab_put_obj(cachep
, slabp
, objp
, node
);
3717 STATS_DEC_ACTIVE(cachep
);
3719 check_slabp(cachep
, slabp
);
3721 /* fixup slab chains */
3722 if (slabp
->inuse
== 0) {
3723 if (l3
->free_objects
> l3
->free_limit
) {
3724 l3
->free_objects
-= cachep
->num
;
3725 /* No need to drop any previously held
3726 * lock here, even if we have a off-slab slab
3727 * descriptor it is guaranteed to come from
3728 * a different cache, refer to comments before
3731 slab_destroy(cachep
, slabp
);
3733 list_add(&slabp
->list
, &l3
->slabs_free
);
3736 /* Unconditionally move a slab to the end of the
3737 * partial list on free - maximum time for the
3738 * other objects to be freed, too.
3740 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3745 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3748 struct kmem_list3
*l3
;
3749 int node
= numa_mem_id();
3751 batchcount
= ac
->batchcount
;
3753 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3756 l3
= cachep
->nodelists
[node
];
3757 spin_lock(&l3
->list_lock
);
3759 struct array_cache
*shared_array
= l3
->shared
;
3760 int max
= shared_array
->limit
- shared_array
->avail
;
3762 if (batchcount
> max
)
3764 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3765 ac
->entry
, sizeof(void *) * batchcount
);
3766 shared_array
->avail
+= batchcount
;
3771 free_block(cachep
, ac
->entry
, batchcount
, node
);
3776 struct list_head
*p
;
3778 p
= l3
->slabs_free
.next
;
3779 while (p
!= &(l3
->slabs_free
)) {
3782 slabp
= list_entry(p
, struct slab
, list
);
3783 BUG_ON(slabp
->inuse
);
3788 STATS_SET_FREEABLE(cachep
, i
);
3791 spin_unlock(&l3
->list_lock
);
3792 ac
->avail
-= batchcount
;
3793 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3797 * Release an obj back to its cache. If the obj has a constructed state, it must
3798 * be in this state _before_ it is released. Called with disabled ints.
3800 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3803 struct array_cache
*ac
= cpu_cache_get(cachep
);
3806 kmemleak_free_recursive(objp
, cachep
->flags
);
3807 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3809 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3812 * Skip calling cache_free_alien() when the platform is not numa.
3813 * This will avoid cache misses that happen while accessing slabp (which
3814 * is per page memory reference) to get nodeid. Instead use a global
3815 * variable to skip the call, which is mostly likely to be present in
3818 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3821 if (likely(ac
->avail
< ac
->limit
)) {
3822 STATS_INC_FREEHIT(cachep
);
3824 STATS_INC_FREEMISS(cachep
);
3825 cache_flusharray(cachep
, ac
);
3828 ac_put_obj(cachep
, ac
, objp
);
3832 * kmem_cache_alloc - Allocate an object
3833 * @cachep: The cache to allocate from.
3834 * @flags: See kmalloc().
3836 * Allocate an object from this cache. The flags are only relevant
3837 * if the cache has no available objects.
3839 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3841 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3843 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3844 cachep
->object_size
, cachep
->size
, flags
);
3848 EXPORT_SYMBOL(kmem_cache_alloc
);
3850 #ifdef CONFIG_TRACING
3852 kmem_cache_alloc_trace(size_t size
, struct kmem_cache
*cachep
, gfp_t flags
)
3856 ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3858 trace_kmalloc(_RET_IP_
, ret
,
3859 size
, slab_buffer_size(cachep
), flags
);
3862 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3866 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3868 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3869 __builtin_return_address(0));
3871 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3872 cachep
->object_size
, cachep
->size
,
3877 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3879 #ifdef CONFIG_TRACING
3880 void *kmem_cache_alloc_node_trace(size_t size
,
3881 struct kmem_cache
*cachep
,
3887 ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3888 __builtin_return_address(0));
3889 trace_kmalloc_node(_RET_IP_
, ret
,
3890 size
, slab_buffer_size(cachep
),
3894 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3897 static __always_inline
void *
3898 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3900 struct kmem_cache
*cachep
;
3902 cachep
= kmem_find_general_cachep(size
, flags
);
3903 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3905 return kmem_cache_alloc_node_trace(size
, cachep
, flags
, node
);
3908 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3909 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3911 return __do_kmalloc_node(size
, flags
, node
,
3912 __builtin_return_address(0));
3914 EXPORT_SYMBOL(__kmalloc_node
);
3916 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3917 int node
, unsigned long caller
)
3919 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3921 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3923 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3925 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3927 EXPORT_SYMBOL(__kmalloc_node
);
3928 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3929 #endif /* CONFIG_NUMA */
3932 * __do_kmalloc - allocate memory
3933 * @size: how many bytes of memory are required.
3934 * @flags: the type of memory to allocate (see kmalloc).
3935 * @caller: function caller for debug tracking of the caller
3937 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3940 struct kmem_cache
*cachep
;
3943 /* If you want to save a few bytes .text space: replace
3945 * Then kmalloc uses the uninlined functions instead of the inline
3948 cachep
= __find_general_cachep(size
, flags
);
3949 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3951 ret
= __cache_alloc(cachep
, flags
, caller
);
3953 trace_kmalloc((unsigned long) caller
, ret
,
3954 size
, cachep
->size
, flags
);
3960 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3961 void *__kmalloc(size_t size
, gfp_t flags
)
3963 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3965 EXPORT_SYMBOL(__kmalloc
);
3967 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3969 return __do_kmalloc(size
, flags
, (void *)caller
);
3971 EXPORT_SYMBOL(__kmalloc_track_caller
);
3974 void *__kmalloc(size_t size
, gfp_t flags
)
3976 return __do_kmalloc(size
, flags
, NULL
);
3978 EXPORT_SYMBOL(__kmalloc
);
3982 * kmem_cache_free - Deallocate an object
3983 * @cachep: The cache the allocation was from.
3984 * @objp: The previously allocated object.
3986 * Free an object which was previously allocated from this
3989 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3991 unsigned long flags
;
3993 local_irq_save(flags
);
3994 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3995 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3996 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3997 __cache_free(cachep
, objp
, __builtin_return_address(0));
3998 local_irq_restore(flags
);
4000 trace_kmem_cache_free(_RET_IP_
, objp
);
4002 EXPORT_SYMBOL(kmem_cache_free
);
4005 * kfree - free previously allocated memory
4006 * @objp: pointer returned by kmalloc.
4008 * If @objp is NULL, no operation is performed.
4010 * Don't free memory not originally allocated by kmalloc()
4011 * or you will run into trouble.
4013 void kfree(const void *objp
)
4015 struct kmem_cache
*c
;
4016 unsigned long flags
;
4018 trace_kfree(_RET_IP_
, objp
);
4020 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
4022 local_irq_save(flags
);
4023 kfree_debugcheck(objp
);
4024 c
= virt_to_cache(objp
);
4025 debug_check_no_locks_freed(objp
, c
->object_size
);
4027 debug_check_no_obj_freed(objp
, c
->object_size
);
4028 __cache_free(c
, (void *)objp
, __builtin_return_address(0));
4029 local_irq_restore(flags
);
4031 EXPORT_SYMBOL(kfree
);
4033 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
4035 return cachep
->object_size
;
4037 EXPORT_SYMBOL(kmem_cache_size
);
4040 * This initializes kmem_list3 or resizes various caches for all nodes.
4042 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
4045 struct kmem_list3
*l3
;
4046 struct array_cache
*new_shared
;
4047 struct array_cache
**new_alien
= NULL
;
4049 for_each_online_node(node
) {
4051 if (use_alien_caches
) {
4052 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
4058 if (cachep
->shared
) {
4059 new_shared
= alloc_arraycache(node
,
4060 cachep
->shared
*cachep
->batchcount
,
4063 free_alien_cache(new_alien
);
4068 l3
= cachep
->nodelists
[node
];
4070 struct array_cache
*shared
= l3
->shared
;
4072 spin_lock_irq(&l3
->list_lock
);
4075 free_block(cachep
, shared
->entry
,
4076 shared
->avail
, node
);
4078 l3
->shared
= new_shared
;
4080 l3
->alien
= new_alien
;
4083 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4084 cachep
->batchcount
+ cachep
->num
;
4085 spin_unlock_irq(&l3
->list_lock
);
4087 free_alien_cache(new_alien
);
4090 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
4092 free_alien_cache(new_alien
);
4097 kmem_list3_init(l3
);
4098 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
4099 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
4100 l3
->shared
= new_shared
;
4101 l3
->alien
= new_alien
;
4102 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
4103 cachep
->batchcount
+ cachep
->num
;
4104 cachep
->nodelists
[node
] = l3
;
4109 if (!cachep
->list
.next
) {
4110 /* Cache is not active yet. Roll back what we did */
4113 if (cachep
->nodelists
[node
]) {
4114 l3
= cachep
->nodelists
[node
];
4117 free_alien_cache(l3
->alien
);
4119 cachep
->nodelists
[node
] = NULL
;
4127 struct ccupdate_struct
{
4128 struct kmem_cache
*cachep
;
4129 struct array_cache
*new[0];
4132 static void do_ccupdate_local(void *info
)
4134 struct ccupdate_struct
*new = info
;
4135 struct array_cache
*old
;
4138 old
= cpu_cache_get(new->cachep
);
4140 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
4141 new->new[smp_processor_id()] = old
;
4144 /* Always called with the slab_mutex held */
4145 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
4146 int batchcount
, int shared
, gfp_t gfp
)
4148 struct ccupdate_struct
*new;
4151 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
4156 for_each_online_cpu(i
) {
4157 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
4160 for (i
--; i
>= 0; i
--)
4166 new->cachep
= cachep
;
4168 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
4171 cachep
->batchcount
= batchcount
;
4172 cachep
->limit
= limit
;
4173 cachep
->shared
= shared
;
4175 for_each_online_cpu(i
) {
4176 struct array_cache
*ccold
= new->new[i
];
4179 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4180 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4181 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4185 return alloc_kmemlist(cachep
, gfp
);
4188 /* Called with slab_mutex held always */
4189 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4195 * The head array serves three purposes:
4196 * - create a LIFO ordering, i.e. return objects that are cache-warm
4197 * - reduce the number of spinlock operations.
4198 * - reduce the number of linked list operations on the slab and
4199 * bufctl chains: array operations are cheaper.
4200 * The numbers are guessed, we should auto-tune as described by
4203 if (cachep
->size
> 131072)
4205 else if (cachep
->size
> PAGE_SIZE
)
4207 else if (cachep
->size
> 1024)
4209 else if (cachep
->size
> 256)
4215 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4216 * allocation behaviour: Most allocs on one cpu, most free operations
4217 * on another cpu. For these cases, an efficient object passing between
4218 * cpus is necessary. This is provided by a shared array. The array
4219 * replaces Bonwick's magazine layer.
4220 * On uniprocessor, it's functionally equivalent (but less efficient)
4221 * to a larger limit. Thus disabled by default.
4224 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4229 * With debugging enabled, large batchcount lead to excessively long
4230 * periods with disabled local interrupts. Limit the batchcount
4235 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4237 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4238 cachep
->name
, -err
);
4243 * Drain an array if it contains any elements taking the l3 lock only if
4244 * necessary. Note that the l3 listlock also protects the array_cache
4245 * if drain_array() is used on the shared array.
4247 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4248 struct array_cache
*ac
, int force
, int node
)
4252 if (!ac
|| !ac
->avail
)
4254 if (ac
->touched
&& !force
) {
4257 spin_lock_irq(&l3
->list_lock
);
4259 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4260 if (tofree
> ac
->avail
)
4261 tofree
= (ac
->avail
+ 1) / 2;
4262 free_block(cachep
, ac
->entry
, tofree
, node
);
4263 ac
->avail
-= tofree
;
4264 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4265 sizeof(void *) * ac
->avail
);
4267 spin_unlock_irq(&l3
->list_lock
);
4272 * cache_reap - Reclaim memory from caches.
4273 * @w: work descriptor
4275 * Called from workqueue/eventd every few seconds.
4277 * - clear the per-cpu caches for this CPU.
4278 * - return freeable pages to the main free memory pool.
4280 * If we cannot acquire the cache chain mutex then just give up - we'll try
4281 * again on the next iteration.
4283 static void cache_reap(struct work_struct
*w
)
4285 struct kmem_cache
*searchp
;
4286 struct kmem_list3
*l3
;
4287 int node
= numa_mem_id();
4288 struct delayed_work
*work
= to_delayed_work(w
);
4290 if (!mutex_trylock(&slab_mutex
))
4291 /* Give up. Setup the next iteration. */
4294 list_for_each_entry(searchp
, &slab_caches
, list
) {
4298 * We only take the l3 lock if absolutely necessary and we
4299 * have established with reasonable certainty that
4300 * we can do some work if the lock was obtained.
4302 l3
= searchp
->nodelists
[node
];
4304 reap_alien(searchp
, l3
);
4306 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4309 * These are racy checks but it does not matter
4310 * if we skip one check or scan twice.
4312 if (time_after(l3
->next_reap
, jiffies
))
4315 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4317 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4319 if (l3
->free_touched
)
4320 l3
->free_touched
= 0;
4324 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4325 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4326 STATS_ADD_REAPED(searchp
, freed
);
4332 mutex_unlock(&slab_mutex
);
4335 /* Set up the next iteration */
4336 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4339 #ifdef CONFIG_SLABINFO
4341 static void print_slabinfo_header(struct seq_file
*m
)
4344 * Output format version, so at least we can change it
4345 * without _too_ many complaints.
4348 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4350 seq_puts(m
, "slabinfo - version: 2.1\n");
4352 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4353 "<objperslab> <pagesperslab>");
4354 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4355 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4357 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4358 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4359 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4364 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4368 mutex_lock(&slab_mutex
);
4370 print_slabinfo_header(m
);
4372 return seq_list_start(&slab_caches
, *pos
);
4375 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4377 return seq_list_next(p
, &slab_caches
, pos
);
4380 static void s_stop(struct seq_file
*m
, void *p
)
4382 mutex_unlock(&slab_mutex
);
4385 static int s_show(struct seq_file
*m
, void *p
)
4387 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4389 unsigned long active_objs
;
4390 unsigned long num_objs
;
4391 unsigned long active_slabs
= 0;
4392 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4396 struct kmem_list3
*l3
;
4400 for_each_online_node(node
) {
4401 l3
= cachep
->nodelists
[node
];
4406 spin_lock_irq(&l3
->list_lock
);
4408 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4409 if (slabp
->inuse
!= cachep
->num
&& !error
)
4410 error
= "slabs_full accounting error";
4411 active_objs
+= cachep
->num
;
4414 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4415 if (slabp
->inuse
== cachep
->num
&& !error
)
4416 error
= "slabs_partial inuse accounting error";
4417 if (!slabp
->inuse
&& !error
)
4418 error
= "slabs_partial/inuse accounting error";
4419 active_objs
+= slabp
->inuse
;
4422 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4423 if (slabp
->inuse
&& !error
)
4424 error
= "slabs_free/inuse accounting error";
4427 free_objects
+= l3
->free_objects
;
4429 shared_avail
+= l3
->shared
->avail
;
4431 spin_unlock_irq(&l3
->list_lock
);
4433 num_slabs
+= active_slabs
;
4434 num_objs
= num_slabs
* cachep
->num
;
4435 if (num_objs
- active_objs
!= free_objects
&& !error
)
4436 error
= "free_objects accounting error";
4438 name
= cachep
->name
;
4440 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4442 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4443 name
, active_objs
, num_objs
, cachep
->size
,
4444 cachep
->num
, (1 << cachep
->gfporder
));
4445 seq_printf(m
, " : tunables %4u %4u %4u",
4446 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4447 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4448 active_slabs
, num_slabs
, shared_avail
);
4451 unsigned long high
= cachep
->high_mark
;
4452 unsigned long allocs
= cachep
->num_allocations
;
4453 unsigned long grown
= cachep
->grown
;
4454 unsigned long reaped
= cachep
->reaped
;
4455 unsigned long errors
= cachep
->errors
;
4456 unsigned long max_freeable
= cachep
->max_freeable
;
4457 unsigned long node_allocs
= cachep
->node_allocs
;
4458 unsigned long node_frees
= cachep
->node_frees
;
4459 unsigned long overflows
= cachep
->node_overflow
;
4461 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4462 "%4lu %4lu %4lu %4lu %4lu",
4463 allocs
, high
, grown
,
4464 reaped
, errors
, max_freeable
, node_allocs
,
4465 node_frees
, overflows
);
4469 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4470 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4471 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4472 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4474 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4475 allochit
, allocmiss
, freehit
, freemiss
);
4483 * slabinfo_op - iterator that generates /proc/slabinfo
4492 * num-pages-per-slab
4493 * + further values on SMP and with statistics enabled
4496 static const struct seq_operations slabinfo_op
= {
4503 #define MAX_SLABINFO_WRITE 128
4505 * slabinfo_write - Tuning for the slab allocator
4507 * @buffer: user buffer
4508 * @count: data length
4511 static ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4512 size_t count
, loff_t
*ppos
)
4514 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4515 int limit
, batchcount
, shared
, res
;
4516 struct kmem_cache
*cachep
;
4518 if (count
> MAX_SLABINFO_WRITE
)
4520 if (copy_from_user(&kbuf
, buffer
, count
))
4522 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4524 tmp
= strchr(kbuf
, ' ');
4529 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4532 /* Find the cache in the chain of caches. */
4533 mutex_lock(&slab_mutex
);
4535 list_for_each_entry(cachep
, &slab_caches
, list
) {
4536 if (!strcmp(cachep
->name
, kbuf
)) {
4537 if (limit
< 1 || batchcount
< 1 ||
4538 batchcount
> limit
|| shared
< 0) {
4541 res
= do_tune_cpucache(cachep
, limit
,
4548 mutex_unlock(&slab_mutex
);
4554 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4556 return seq_open(file
, &slabinfo_op
);
4559 static const struct file_operations proc_slabinfo_operations
= {
4560 .open
= slabinfo_open
,
4562 .write
= slabinfo_write
,
4563 .llseek
= seq_lseek
,
4564 .release
= seq_release
,
4567 #ifdef CONFIG_DEBUG_SLAB_LEAK
4569 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4571 mutex_lock(&slab_mutex
);
4572 return seq_list_start(&slab_caches
, *pos
);
4575 static inline int add_caller(unsigned long *n
, unsigned long v
)
4585 unsigned long *q
= p
+ 2 * i
;
4599 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4605 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4611 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4612 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4614 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4619 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4621 #ifdef CONFIG_KALLSYMS
4622 unsigned long offset
, size
;
4623 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4625 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4626 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4628 seq_printf(m
, " [%s]", modname
);
4632 seq_printf(m
, "%p", (void *)address
);
4635 static int leaks_show(struct seq_file
*m
, void *p
)
4637 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4639 struct kmem_list3
*l3
;
4641 unsigned long *n
= m
->private;
4645 if (!(cachep
->flags
& SLAB_STORE_USER
))
4647 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4650 /* OK, we can do it */
4654 for_each_online_node(node
) {
4655 l3
= cachep
->nodelists
[node
];
4660 spin_lock_irq(&l3
->list_lock
);
4662 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4663 handle_slab(n
, cachep
, slabp
);
4664 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4665 handle_slab(n
, cachep
, slabp
);
4666 spin_unlock_irq(&l3
->list_lock
);
4668 name
= cachep
->name
;
4670 /* Increase the buffer size */
4671 mutex_unlock(&slab_mutex
);
4672 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4674 /* Too bad, we are really out */
4676 mutex_lock(&slab_mutex
);
4679 *(unsigned long *)m
->private = n
[0] * 2;
4681 mutex_lock(&slab_mutex
);
4682 /* Now make sure this entry will be retried */
4686 for (i
= 0; i
< n
[1]; i
++) {
4687 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4688 show_symbol(m
, n
[2*i
+2]);
4695 static const struct seq_operations slabstats_op
= {
4696 .start
= leaks_start
,
4702 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4704 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4707 ret
= seq_open(file
, &slabstats_op
);
4709 struct seq_file
*m
= file
->private_data
;
4710 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4719 static const struct file_operations proc_slabstats_operations
= {
4720 .open
= slabstats_open
,
4722 .llseek
= seq_lseek
,
4723 .release
= seq_release_private
,
4727 static int __init
slab_proc_init(void)
4729 proc_create("slabinfo",S_IWUSR
|S_IRUSR
,NULL
,&proc_slabinfo_operations
);
4730 #ifdef CONFIG_DEBUG_SLAB_LEAK
4731 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4735 module_init(slab_proc_init
);
4739 * ksize - get the actual amount of memory allocated for a given object
4740 * @objp: Pointer to the object
4742 * kmalloc may internally round up allocations and return more memory
4743 * than requested. ksize() can be used to determine the actual amount of
4744 * memory allocated. The caller may use this additional memory, even though
4745 * a smaller amount of memory was initially specified with the kmalloc call.
4746 * The caller must guarantee that objp points to a valid object previously
4747 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4748 * must not be freed during the duration of the call.
4750 size_t ksize(const void *objp
)
4753 if (unlikely(objp
== ZERO_SIZE_PTR
))
4756 return virt_to_cache(objp
)->object_size
;
4758 EXPORT_SYMBOL(ksize
);