3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
121 #include <asm/cacheflush.h>
122 #include <asm/tlbflush.h>
123 #include <asm/page.h>
125 #include <trace/events/kmem.h>
128 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
129 * 0 for faster, smaller code (especially in the critical paths).
131 * STATS - 1 to collect stats for /proc/slabinfo.
132 * 0 for faster, smaller code (especially in the critical paths).
134 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
137 #ifdef CONFIG_DEBUG_SLAB
140 #define FORCED_DEBUG 1
144 #define FORCED_DEBUG 0
147 /* Shouldn't this be in a header file somewhere? */
148 #define BYTES_PER_WORD sizeof(void *)
149 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
151 #ifndef ARCH_KMALLOC_FLAGS
152 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
155 /* Legal flag mask for kmem_cache_create(). */
157 # define CREATE_MASK (SLAB_RED_ZONE | \
158 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
161 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
162 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
163 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
165 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
167 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
168 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
169 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
175 * Bufctl's are used for linking objs within a slab
178 * This implementation relies on "struct page" for locating the cache &
179 * slab an object belongs to.
180 * This allows the bufctl structure to be small (one int), but limits
181 * the number of objects a slab (not a cache) can contain when off-slab
182 * bufctls are used. The limit is the size of the largest general cache
183 * that does not use off-slab slabs.
184 * For 32bit archs with 4 kB pages, is this 56.
185 * This is not serious, as it is only for large objects, when it is unwise
186 * to have too many per slab.
187 * Note: This limit can be raised by introducing a general cache whose size
188 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
191 typedef unsigned int kmem_bufctl_t
;
192 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
193 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
194 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
195 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
200 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
201 * arrange for kmem_freepages to be called via RCU. This is useful if
202 * we need to approach a kernel structure obliquely, from its address
203 * obtained without the usual locking. We can lock the structure to
204 * stabilize it and check it's still at the given address, only if we
205 * can be sure that the memory has not been meanwhile reused for some
206 * other kind of object (which our subsystem's lock might corrupt).
208 * rcu_read_lock before reading the address, then rcu_read_unlock after
209 * taking the spinlock within the structure expected at that address.
212 struct rcu_head head
;
213 struct kmem_cache
*cachep
;
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
227 struct list_head list
;
228 unsigned long colouroff
;
229 void *s_mem
; /* including colour offset */
230 unsigned int inuse
; /* num of objs active in slab */
232 unsigned short nodeid
;
234 struct slab_rcu __slab_cover_slab_rcu
;
242 * - LIFO ordering, to hand out cache-warm objects from _alloc
243 * - reduce the number of linked list operations
244 * - reduce spinlock operations
246 * The limit is stored in the per-cpu structure to reduce the data cache
253 unsigned int batchcount
;
254 unsigned int touched
;
257 * Must have this definition in here for the proper
258 * alignment of array_cache. Also simplifies accessing
264 * bootstrap: The caches do not work without cpuarrays anymore, but the
265 * cpuarrays are allocated from the generic caches...
267 #define BOOT_CPUCACHE_ENTRIES 1
268 struct arraycache_init
{
269 struct array_cache cache
;
270 void *entries
[BOOT_CPUCACHE_ENTRIES
];
274 * The slab lists for all objects.
277 struct list_head slabs_partial
; /* partial list first, better asm code */
278 struct list_head slabs_full
;
279 struct list_head slabs_free
;
280 unsigned long free_objects
;
281 unsigned int free_limit
;
282 unsigned int colour_next
; /* Per-node cache coloring */
283 spinlock_t list_lock
;
284 struct array_cache
*shared
; /* shared per node */
285 struct array_cache
**alien
; /* on other nodes */
286 unsigned long next_reap
; /* updated without locking */
287 int free_touched
; /* updated without locking */
291 * Need this for bootstrapping a per node allocator.
293 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
294 static struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
295 #define CACHE_CACHE 0
296 #define SIZE_AC MAX_NUMNODES
297 #define SIZE_L3 (2 * MAX_NUMNODES)
299 static int drain_freelist(struct kmem_cache
*cache
,
300 struct kmem_list3
*l3
, int tofree
);
301 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
303 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
304 static void cache_reap(struct work_struct
*unused
);
307 * This function must be completely optimized away if a constant is passed to
308 * it. Mostly the same as what is in linux/slab.h except it returns an index.
310 static __always_inline
int index_of(const size_t size
)
312 extern void __bad_size(void);
314 if (__builtin_constant_p(size
)) {
322 #include <linux/kmalloc_sizes.h>
330 static int slab_early_init
= 1;
332 #define INDEX_AC index_of(sizeof(struct arraycache_init))
333 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
335 static void kmem_list3_init(struct kmem_list3
*parent
)
337 INIT_LIST_HEAD(&parent
->slabs_full
);
338 INIT_LIST_HEAD(&parent
->slabs_partial
);
339 INIT_LIST_HEAD(&parent
->slabs_free
);
340 parent
->shared
= NULL
;
341 parent
->alien
= NULL
;
342 parent
->colour_next
= 0;
343 spin_lock_init(&parent
->list_lock
);
344 parent
->free_objects
= 0;
345 parent
->free_touched
= 0;
348 #define MAKE_LIST(cachep, listp, slab, nodeid) \
350 INIT_LIST_HEAD(listp); \
351 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
354 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
356 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
358 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
361 #define CFLGS_OFF_SLAB (0x80000000UL)
362 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
364 #define BATCHREFILL_LIMIT 16
366 * Optimization question: fewer reaps means less probability for unnessary
367 * cpucache drain/refill cycles.
369 * OTOH the cpuarrays can contain lots of objects,
370 * which could lock up otherwise freeable slabs.
372 #define REAPTIMEOUT_CPUC (2*HZ)
373 #define REAPTIMEOUT_LIST3 (4*HZ)
376 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
377 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
378 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
379 #define STATS_INC_GROWN(x) ((x)->grown++)
380 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
381 #define STATS_SET_HIGH(x) \
383 if ((x)->num_active > (x)->high_mark) \
384 (x)->high_mark = (x)->num_active; \
386 #define STATS_INC_ERR(x) ((x)->errors++)
387 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
388 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
389 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
390 #define STATS_SET_FREEABLE(x, i) \
392 if ((x)->max_freeable < i) \
393 (x)->max_freeable = i; \
395 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
396 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
397 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
398 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
400 #define STATS_INC_ACTIVE(x) do { } while (0)
401 #define STATS_DEC_ACTIVE(x) do { } while (0)
402 #define STATS_INC_ALLOCED(x) do { } while (0)
403 #define STATS_INC_GROWN(x) do { } while (0)
404 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
405 #define STATS_SET_HIGH(x) do { } while (0)
406 #define STATS_INC_ERR(x) do { } while (0)
407 #define STATS_INC_NODEALLOCS(x) do { } while (0)
408 #define STATS_INC_NODEFREES(x) do { } while (0)
409 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
410 #define STATS_SET_FREEABLE(x, i) do { } while (0)
411 #define STATS_INC_ALLOCHIT(x) do { } while (0)
412 #define STATS_INC_ALLOCMISS(x) do { } while (0)
413 #define STATS_INC_FREEHIT(x) do { } while (0)
414 #define STATS_INC_FREEMISS(x) do { } while (0)
420 * memory layout of objects:
422 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
423 * the end of an object is aligned with the end of the real
424 * allocation. Catches writes behind the end of the allocation.
425 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
427 * cachep->obj_offset: The real object.
428 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
429 * cachep->size - 1* BYTES_PER_WORD: last caller address
430 * [BYTES_PER_WORD long]
432 static int obj_offset(struct kmem_cache
*cachep
)
434 return cachep
->obj_offset
;
437 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
439 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
440 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
441 sizeof(unsigned long long));
444 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
446 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
447 if (cachep
->flags
& SLAB_STORE_USER
)
448 return (unsigned long long *)(objp
+ cachep
->size
-
449 sizeof(unsigned long long) -
451 return (unsigned long long *) (objp
+ cachep
->size
-
452 sizeof(unsigned long long));
455 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
457 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
458 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
463 #define obj_offset(x) 0
464 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
466 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470 #ifdef CONFIG_TRACING
471 size_t slab_buffer_size(struct kmem_cache
*cachep
)
475 EXPORT_SYMBOL(slab_buffer_size
);
479 * Do not go above this order unless 0 objects fit into the slab or
480 * overridden on the command line.
482 #define SLAB_MAX_ORDER_HI 1
483 #define SLAB_MAX_ORDER_LO 0
484 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
485 static bool slab_max_order_set __initdata
;
487 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
489 page
= compound_head(page
);
490 BUG_ON(!PageSlab(page
));
491 return page
->slab_cache
;
494 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
496 struct page
*page
= virt_to_head_page(obj
);
497 return page
->slab_cache
;
500 static inline struct slab
*virt_to_slab(const void *obj
)
502 struct page
*page
= virt_to_head_page(obj
);
504 VM_BUG_ON(!PageSlab(page
));
505 return page
->slab_page
;
508 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
511 return slab
->s_mem
+ cache
->size
* idx
;
515 * We want to avoid an expensive divide : (offset / cache->size)
516 * Using the fact that size is a constant for a particular cache,
517 * we can replace (offset / cache->size) by
518 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
520 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
521 const struct slab
*slab
, void *obj
)
523 u32 offset
= (obj
- slab
->s_mem
);
524 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
528 * These are the default caches for kmalloc. Custom caches can have other sizes.
530 struct cache_sizes malloc_sizes
[] = {
531 #define CACHE(x) { .cs_size = (x) },
532 #include <linux/kmalloc_sizes.h>
536 EXPORT_SYMBOL(malloc_sizes
);
538 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
544 static struct cache_names __initdata cache_names
[] = {
545 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
546 #include <linux/kmalloc_sizes.h>
551 static struct arraycache_init initarray_cache __initdata
=
552 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
553 static struct arraycache_init initarray_generic
=
554 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
556 /* internal cache of cache description objs */
557 static struct kmem_list3
*cache_cache_nodelists
[MAX_NUMNODES
];
558 static struct kmem_cache cache_cache
= {
559 .nodelists
= cache_cache_nodelists
,
561 .limit
= BOOT_CPUCACHE_ENTRIES
,
563 .size
= sizeof(struct kmem_cache
),
564 .name
= "kmem_cache",
567 #define BAD_ALIEN_MAGIC 0x01020304ul
569 #ifdef CONFIG_LOCKDEP
572 * Slab sometimes uses the kmalloc slabs to store the slab headers
573 * for other slabs "off slab".
574 * The locking for this is tricky in that it nests within the locks
575 * of all other slabs in a few places; to deal with this special
576 * locking we put on-slab caches into a separate lock-class.
578 * We set lock class for alien array caches which are up during init.
579 * The lock annotation will be lost if all cpus of a node goes down and
580 * then comes back up during hotplug
582 static struct lock_class_key on_slab_l3_key
;
583 static struct lock_class_key on_slab_alc_key
;
585 static struct lock_class_key debugobj_l3_key
;
586 static struct lock_class_key debugobj_alc_key
;
588 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
589 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
592 struct array_cache
**alc
;
593 struct kmem_list3
*l3
;
596 l3
= cachep
->nodelists
[q
];
600 lockdep_set_class(&l3
->list_lock
, l3_key
);
603 * FIXME: This check for BAD_ALIEN_MAGIC
604 * should go away when common slab code is taught to
605 * work even without alien caches.
606 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
607 * for alloc_alien_cache,
609 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
613 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
617 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
619 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
622 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
626 for_each_online_node(node
)
627 slab_set_debugobj_lock_classes_node(cachep
, node
);
630 static void init_node_lock_keys(int q
)
632 struct cache_sizes
*s
= malloc_sizes
;
637 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
638 struct kmem_list3
*l3
;
640 l3
= s
->cs_cachep
->nodelists
[q
];
641 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
644 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
645 &on_slab_alc_key
, q
);
649 static inline void init_lock_keys(void)
654 init_node_lock_keys(node
);
657 static void init_node_lock_keys(int q
)
661 static inline void init_lock_keys(void)
665 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
669 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
674 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
676 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
678 return cachep
->array
[smp_processor_id()];
681 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
684 struct cache_sizes
*csizep
= malloc_sizes
;
687 /* This happens if someone tries to call
688 * kmem_cache_create(), or __kmalloc(), before
689 * the generic caches are initialized.
691 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
694 return ZERO_SIZE_PTR
;
696 while (size
> csizep
->cs_size
)
700 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
701 * has cs_{dma,}cachep==NULL. Thus no special case
702 * for large kmalloc calls required.
704 #ifdef CONFIG_ZONE_DMA
705 if (unlikely(gfpflags
& GFP_DMA
))
706 return csizep
->cs_dmacachep
;
708 return csizep
->cs_cachep
;
711 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
713 return __find_general_cachep(size
, gfpflags
);
716 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
718 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
722 * Calculate the number of objects and left-over bytes for a given buffer size.
724 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
725 size_t align
, int flags
, size_t *left_over
,
730 size_t slab_size
= PAGE_SIZE
<< gfporder
;
733 * The slab management structure can be either off the slab or
734 * on it. For the latter case, the memory allocated for a
738 * - One kmem_bufctl_t for each object
739 * - Padding to respect alignment of @align
740 * - @buffer_size bytes for each object
742 * If the slab management structure is off the slab, then the
743 * alignment will already be calculated into the size. Because
744 * the slabs are all pages aligned, the objects will be at the
745 * correct alignment when allocated.
747 if (flags
& CFLGS_OFF_SLAB
) {
749 nr_objs
= slab_size
/ buffer_size
;
751 if (nr_objs
> SLAB_LIMIT
)
752 nr_objs
= SLAB_LIMIT
;
755 * Ignore padding for the initial guess. The padding
756 * is at most @align-1 bytes, and @buffer_size is at
757 * least @align. In the worst case, this result will
758 * be one greater than the number of objects that fit
759 * into the memory allocation when taking the padding
762 nr_objs
= (slab_size
- sizeof(struct slab
)) /
763 (buffer_size
+ sizeof(kmem_bufctl_t
));
766 * This calculated number will be either the right
767 * amount, or one greater than what we want.
769 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
773 if (nr_objs
> SLAB_LIMIT
)
774 nr_objs
= SLAB_LIMIT
;
776 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
779 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
782 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
784 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
787 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
788 function
, cachep
->name
, msg
);
793 * By default on NUMA we use alien caches to stage the freeing of
794 * objects allocated from other nodes. This causes massive memory
795 * inefficiencies when using fake NUMA setup to split memory into a
796 * large number of small nodes, so it can be disabled on the command
800 static int use_alien_caches __read_mostly
= 1;
801 static int __init
noaliencache_setup(char *s
)
803 use_alien_caches
= 0;
806 __setup("noaliencache", noaliencache_setup
);
808 static int __init
slab_max_order_setup(char *str
)
810 get_option(&str
, &slab_max_order
);
811 slab_max_order
= slab_max_order
< 0 ? 0 :
812 min(slab_max_order
, MAX_ORDER
- 1);
813 slab_max_order_set
= true;
817 __setup("slab_max_order=", slab_max_order_setup
);
821 * Special reaping functions for NUMA systems called from cache_reap().
822 * These take care of doing round robin flushing of alien caches (containing
823 * objects freed on different nodes from which they were allocated) and the
824 * flushing of remote pcps by calling drain_node_pages.
826 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
828 static void init_reap_node(int cpu
)
832 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
833 if (node
== MAX_NUMNODES
)
834 node
= first_node(node_online_map
);
836 per_cpu(slab_reap_node
, cpu
) = node
;
839 static void next_reap_node(void)
841 int node
= __this_cpu_read(slab_reap_node
);
843 node
= next_node(node
, node_online_map
);
844 if (unlikely(node
>= MAX_NUMNODES
))
845 node
= first_node(node_online_map
);
846 __this_cpu_write(slab_reap_node
, node
);
850 #define init_reap_node(cpu) do { } while (0)
851 #define next_reap_node(void) do { } while (0)
855 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
856 * via the workqueue/eventd.
857 * Add the CPU number into the expiration time to minimize the possibility of
858 * the CPUs getting into lockstep and contending for the global cache chain
861 static void __cpuinit
start_cpu_timer(int cpu
)
863 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
866 * When this gets called from do_initcalls via cpucache_init(),
867 * init_workqueues() has already run, so keventd will be setup
870 if (keventd_up() && reap_work
->work
.func
== NULL
) {
872 INIT_DELAYED_WORK_DEFERRABLE(reap_work
, cache_reap
);
873 schedule_delayed_work_on(cpu
, reap_work
,
874 __round_jiffies_relative(HZ
, cpu
));
878 static struct array_cache
*alloc_arraycache(int node
, int entries
,
879 int batchcount
, gfp_t gfp
)
881 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
882 struct array_cache
*nc
= NULL
;
884 nc
= kmalloc_node(memsize
, gfp
, node
);
886 * The array_cache structures contain pointers to free object.
887 * However, when such objects are allocated or transferred to another
888 * cache the pointers are not cleared and they could be counted as
889 * valid references during a kmemleak scan. Therefore, kmemleak must
890 * not scan such objects.
892 kmemleak_no_scan(nc
);
896 nc
->batchcount
= batchcount
;
898 spin_lock_init(&nc
->lock
);
904 * Transfer objects in one arraycache to another.
905 * Locking must be handled by the caller.
907 * Return the number of entries transferred.
909 static int transfer_objects(struct array_cache
*to
,
910 struct array_cache
*from
, unsigned int max
)
912 /* Figure out how many entries to transfer */
913 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
918 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
928 #define drain_alien_cache(cachep, alien) do { } while (0)
929 #define reap_alien(cachep, l3) do { } while (0)
931 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
933 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
936 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
940 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
945 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
951 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
952 gfp_t flags
, int nodeid
)
957 #else /* CONFIG_NUMA */
959 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
960 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
962 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
964 struct array_cache
**ac_ptr
;
965 int memsize
= sizeof(void *) * nr_node_ids
;
970 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
973 if (i
== node
|| !node_online(i
))
975 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
977 for (i
--; i
>= 0; i
--)
987 static void free_alien_cache(struct array_cache
**ac_ptr
)
998 static void __drain_alien_cache(struct kmem_cache
*cachep
,
999 struct array_cache
*ac
, int node
)
1001 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1004 spin_lock(&rl3
->list_lock
);
1006 * Stuff objects into the remote nodes shared array first.
1007 * That way we could avoid the overhead of putting the objects
1008 * into the free lists and getting them back later.
1011 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1013 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1015 spin_unlock(&rl3
->list_lock
);
1020 * Called from cache_reap() to regularly drain alien caches round robin.
1022 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1024 int node
= __this_cpu_read(slab_reap_node
);
1027 struct array_cache
*ac
= l3
->alien
[node
];
1029 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1030 __drain_alien_cache(cachep
, ac
, node
);
1031 spin_unlock_irq(&ac
->lock
);
1036 static void drain_alien_cache(struct kmem_cache
*cachep
,
1037 struct array_cache
**alien
)
1040 struct array_cache
*ac
;
1041 unsigned long flags
;
1043 for_each_online_node(i
) {
1046 spin_lock_irqsave(&ac
->lock
, flags
);
1047 __drain_alien_cache(cachep
, ac
, i
);
1048 spin_unlock_irqrestore(&ac
->lock
, flags
);
1053 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1055 struct slab
*slabp
= virt_to_slab(objp
);
1056 int nodeid
= slabp
->nodeid
;
1057 struct kmem_list3
*l3
;
1058 struct array_cache
*alien
= NULL
;
1061 node
= numa_mem_id();
1064 * Make sure we are not freeing a object from another node to the array
1065 * cache on this cpu.
1067 if (likely(slabp
->nodeid
== node
))
1070 l3
= cachep
->nodelists
[node
];
1071 STATS_INC_NODEFREES(cachep
);
1072 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1073 alien
= l3
->alien
[nodeid
];
1074 spin_lock(&alien
->lock
);
1075 if (unlikely(alien
->avail
== alien
->limit
)) {
1076 STATS_INC_ACOVERFLOW(cachep
);
1077 __drain_alien_cache(cachep
, alien
, nodeid
);
1079 alien
->entry
[alien
->avail
++] = objp
;
1080 spin_unlock(&alien
->lock
);
1082 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1083 free_block(cachep
, &objp
, 1, nodeid
);
1084 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1091 * Allocates and initializes nodelists for a node on each slab cache, used for
1092 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1093 * will be allocated off-node since memory is not yet online for the new node.
1094 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1097 * Must hold slab_mutex.
1099 static int init_cache_nodelists_node(int node
)
1101 struct kmem_cache
*cachep
;
1102 struct kmem_list3
*l3
;
1103 const int memsize
= sizeof(struct kmem_list3
);
1105 list_for_each_entry(cachep
, &slab_caches
, list
) {
1107 * Set up the size64 kmemlist for cpu before we can
1108 * begin anything. Make sure some other cpu on this
1109 * node has not already allocated this
1111 if (!cachep
->nodelists
[node
]) {
1112 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1115 kmem_list3_init(l3
);
1116 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1117 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1120 * The l3s don't come and go as CPUs come and
1121 * go. slab_mutex is sufficient
1124 cachep
->nodelists
[node
] = l3
;
1127 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1128 cachep
->nodelists
[node
]->free_limit
=
1129 (1 + nr_cpus_node(node
)) *
1130 cachep
->batchcount
+ cachep
->num
;
1131 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1136 static void __cpuinit
cpuup_canceled(long cpu
)
1138 struct kmem_cache
*cachep
;
1139 struct kmem_list3
*l3
= NULL
;
1140 int node
= cpu_to_mem(cpu
);
1141 const struct cpumask
*mask
= cpumask_of_node(node
);
1143 list_for_each_entry(cachep
, &slab_caches
, list
) {
1144 struct array_cache
*nc
;
1145 struct array_cache
*shared
;
1146 struct array_cache
**alien
;
1148 /* cpu is dead; no one can alloc from it. */
1149 nc
= cachep
->array
[cpu
];
1150 cachep
->array
[cpu
] = NULL
;
1151 l3
= cachep
->nodelists
[node
];
1154 goto free_array_cache
;
1156 spin_lock_irq(&l3
->list_lock
);
1158 /* Free limit for this kmem_list3 */
1159 l3
->free_limit
-= cachep
->batchcount
;
1161 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1163 if (!cpumask_empty(mask
)) {
1164 spin_unlock_irq(&l3
->list_lock
);
1165 goto free_array_cache
;
1168 shared
= l3
->shared
;
1170 free_block(cachep
, shared
->entry
,
1171 shared
->avail
, node
);
1178 spin_unlock_irq(&l3
->list_lock
);
1182 drain_alien_cache(cachep
, alien
);
1183 free_alien_cache(alien
);
1189 * In the previous loop, all the objects were freed to
1190 * the respective cache's slabs, now we can go ahead and
1191 * shrink each nodelist to its limit.
1193 list_for_each_entry(cachep
, &slab_caches
, list
) {
1194 l3
= cachep
->nodelists
[node
];
1197 drain_freelist(cachep
, l3
, l3
->free_objects
);
1201 static int __cpuinit
cpuup_prepare(long cpu
)
1203 struct kmem_cache
*cachep
;
1204 struct kmem_list3
*l3
= NULL
;
1205 int node
= cpu_to_mem(cpu
);
1209 * We need to do this right in the beginning since
1210 * alloc_arraycache's are going to use this list.
1211 * kmalloc_node allows us to add the slab to the right
1212 * kmem_list3 and not this cpu's kmem_list3
1214 err
= init_cache_nodelists_node(node
);
1219 * Now we can go ahead with allocating the shared arrays and
1222 list_for_each_entry(cachep
, &slab_caches
, list
) {
1223 struct array_cache
*nc
;
1224 struct array_cache
*shared
= NULL
;
1225 struct array_cache
**alien
= NULL
;
1227 nc
= alloc_arraycache(node
, cachep
->limit
,
1228 cachep
->batchcount
, GFP_KERNEL
);
1231 if (cachep
->shared
) {
1232 shared
= alloc_arraycache(node
,
1233 cachep
->shared
* cachep
->batchcount
,
1234 0xbaadf00d, GFP_KERNEL
);
1240 if (use_alien_caches
) {
1241 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1248 cachep
->array
[cpu
] = nc
;
1249 l3
= cachep
->nodelists
[node
];
1252 spin_lock_irq(&l3
->list_lock
);
1255 * We are serialised from CPU_DEAD or
1256 * CPU_UP_CANCELLED by the cpucontrol lock
1258 l3
->shared
= shared
;
1267 spin_unlock_irq(&l3
->list_lock
);
1269 free_alien_cache(alien
);
1270 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1271 slab_set_debugobj_lock_classes_node(cachep
, node
);
1273 init_node_lock_keys(node
);
1277 cpuup_canceled(cpu
);
1281 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1282 unsigned long action
, void *hcpu
)
1284 long cpu
= (long)hcpu
;
1288 case CPU_UP_PREPARE
:
1289 case CPU_UP_PREPARE_FROZEN
:
1290 mutex_lock(&slab_mutex
);
1291 err
= cpuup_prepare(cpu
);
1292 mutex_unlock(&slab_mutex
);
1295 case CPU_ONLINE_FROZEN
:
1296 start_cpu_timer(cpu
);
1298 #ifdef CONFIG_HOTPLUG_CPU
1299 case CPU_DOWN_PREPARE
:
1300 case CPU_DOWN_PREPARE_FROZEN
:
1302 * Shutdown cache reaper. Note that the slab_mutex is
1303 * held so that if cache_reap() is invoked it cannot do
1304 * anything expensive but will only modify reap_work
1305 * and reschedule the timer.
1307 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1308 /* Now the cache_reaper is guaranteed to be not running. */
1309 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1311 case CPU_DOWN_FAILED
:
1312 case CPU_DOWN_FAILED_FROZEN
:
1313 start_cpu_timer(cpu
);
1316 case CPU_DEAD_FROZEN
:
1318 * Even if all the cpus of a node are down, we don't free the
1319 * kmem_list3 of any cache. This to avoid a race between
1320 * cpu_down, and a kmalloc allocation from another cpu for
1321 * memory from the node of the cpu going down. The list3
1322 * structure is usually allocated from kmem_cache_create() and
1323 * gets destroyed at kmem_cache_destroy().
1327 case CPU_UP_CANCELED
:
1328 case CPU_UP_CANCELED_FROZEN
:
1329 mutex_lock(&slab_mutex
);
1330 cpuup_canceled(cpu
);
1331 mutex_unlock(&slab_mutex
);
1334 return notifier_from_errno(err
);
1337 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1338 &cpuup_callback
, NULL
, 0
1341 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1343 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1344 * Returns -EBUSY if all objects cannot be drained so that the node is not
1347 * Must hold slab_mutex.
1349 static int __meminit
drain_cache_nodelists_node(int node
)
1351 struct kmem_cache
*cachep
;
1354 list_for_each_entry(cachep
, &slab_caches
, list
) {
1355 struct kmem_list3
*l3
;
1357 l3
= cachep
->nodelists
[node
];
1361 drain_freelist(cachep
, l3
, l3
->free_objects
);
1363 if (!list_empty(&l3
->slabs_full
) ||
1364 !list_empty(&l3
->slabs_partial
)) {
1372 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1373 unsigned long action
, void *arg
)
1375 struct memory_notify
*mnb
= arg
;
1379 nid
= mnb
->status_change_nid
;
1384 case MEM_GOING_ONLINE
:
1385 mutex_lock(&slab_mutex
);
1386 ret
= init_cache_nodelists_node(nid
);
1387 mutex_unlock(&slab_mutex
);
1389 case MEM_GOING_OFFLINE
:
1390 mutex_lock(&slab_mutex
);
1391 ret
= drain_cache_nodelists_node(nid
);
1392 mutex_unlock(&slab_mutex
);
1396 case MEM_CANCEL_ONLINE
:
1397 case MEM_CANCEL_OFFLINE
:
1401 return notifier_from_errno(ret
);
1403 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1406 * swap the static kmem_list3 with kmalloced memory
1408 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1411 struct kmem_list3
*ptr
;
1413 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1416 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1418 * Do not assume that spinlocks can be initialized via memcpy:
1420 spin_lock_init(&ptr
->list_lock
);
1422 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1423 cachep
->nodelists
[nodeid
] = ptr
;
1427 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1428 * size of kmem_list3.
1430 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1434 for_each_online_node(node
) {
1435 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1436 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1438 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1443 * Initialisation. Called after the page allocator have been initialised and
1444 * before smp_init().
1446 void __init
kmem_cache_init(void)
1449 struct cache_sizes
*sizes
;
1450 struct cache_names
*names
;
1455 if (num_possible_nodes() == 1)
1456 use_alien_caches
= 0;
1458 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1459 kmem_list3_init(&initkmem_list3
[i
]);
1460 if (i
< MAX_NUMNODES
)
1461 cache_cache
.nodelists
[i
] = NULL
;
1463 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1466 * Fragmentation resistance on low memory - only use bigger
1467 * page orders on machines with more than 32MB of memory if
1468 * not overridden on the command line.
1470 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1471 slab_max_order
= SLAB_MAX_ORDER_HI
;
1473 /* Bootstrap is tricky, because several objects are allocated
1474 * from caches that do not exist yet:
1475 * 1) initialize the cache_cache cache: it contains the struct
1476 * kmem_cache structures of all caches, except cache_cache itself:
1477 * cache_cache is statically allocated.
1478 * Initially an __init data area is used for the head array and the
1479 * kmem_list3 structures, it's replaced with a kmalloc allocated
1480 * array at the end of the bootstrap.
1481 * 2) Create the first kmalloc cache.
1482 * The struct kmem_cache for the new cache is allocated normally.
1483 * An __init data area is used for the head array.
1484 * 3) Create the remaining kmalloc caches, with minimally sized
1486 * 4) Replace the __init data head arrays for cache_cache and the first
1487 * kmalloc cache with kmalloc allocated arrays.
1488 * 5) Replace the __init data for kmem_list3 for cache_cache and
1489 * the other cache's with kmalloc allocated memory.
1490 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1493 node
= numa_mem_id();
1495 /* 1) create the cache_cache */
1496 INIT_LIST_HEAD(&slab_caches
);
1497 list_add(&cache_cache
.list
, &slab_caches
);
1498 cache_cache
.colour_off
= cache_line_size();
1499 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1500 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1503 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1505 cache_cache
.size
= offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1506 nr_node_ids
* sizeof(struct kmem_list3
*);
1507 cache_cache
.object_size
= cache_cache
.size
;
1508 cache_cache
.size
= ALIGN(cache_cache
.size
,
1510 cache_cache
.reciprocal_buffer_size
=
1511 reciprocal_value(cache_cache
.size
);
1513 for (order
= 0; order
< MAX_ORDER
; order
++) {
1514 cache_estimate(order
, cache_cache
.size
,
1515 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1516 if (cache_cache
.num
)
1519 BUG_ON(!cache_cache
.num
);
1520 cache_cache
.gfporder
= order
;
1521 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1522 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1523 sizeof(struct slab
), cache_line_size());
1525 /* 2+3) create the kmalloc caches */
1526 sizes
= malloc_sizes
;
1527 names
= cache_names
;
1530 * Initialize the caches that provide memory for the array cache and the
1531 * kmem_list3 structures first. Without this, further allocations will
1535 sizes
[INDEX_AC
].cs_cachep
= __kmem_cache_create(names
[INDEX_AC
].name
,
1536 sizes
[INDEX_AC
].cs_size
,
1537 ARCH_KMALLOC_MINALIGN
,
1538 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1541 if (INDEX_AC
!= INDEX_L3
) {
1542 sizes
[INDEX_L3
].cs_cachep
=
1543 __kmem_cache_create(names
[INDEX_L3
].name
,
1544 sizes
[INDEX_L3
].cs_size
,
1545 ARCH_KMALLOC_MINALIGN
,
1546 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1550 slab_early_init
= 0;
1552 while (sizes
->cs_size
!= ULONG_MAX
) {
1554 * For performance, all the general caches are L1 aligned.
1555 * This should be particularly beneficial on SMP boxes, as it
1556 * eliminates "false sharing".
1557 * Note for systems short on memory removing the alignment will
1558 * allow tighter packing of the smaller caches.
1560 if (!sizes
->cs_cachep
) {
1561 sizes
->cs_cachep
= __kmem_cache_create(names
->name
,
1563 ARCH_KMALLOC_MINALIGN
,
1564 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1567 #ifdef CONFIG_ZONE_DMA
1568 sizes
->cs_dmacachep
= __kmem_cache_create(
1571 ARCH_KMALLOC_MINALIGN
,
1572 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1579 /* 4) Replace the bootstrap head arrays */
1581 struct array_cache
*ptr
;
1583 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1585 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1586 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1587 sizeof(struct arraycache_init
));
1589 * Do not assume that spinlocks can be initialized via memcpy:
1591 spin_lock_init(&ptr
->lock
);
1593 cache_cache
.array
[smp_processor_id()] = ptr
;
1595 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1597 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1598 != &initarray_generic
.cache
);
1599 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1600 sizeof(struct arraycache_init
));
1602 * Do not assume that spinlocks can be initialized via memcpy:
1604 spin_lock_init(&ptr
->lock
);
1606 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1609 /* 5) Replace the bootstrap kmem_list3's */
1613 for_each_online_node(nid
) {
1614 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1616 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1617 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1619 if (INDEX_AC
!= INDEX_L3
) {
1620 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1621 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1629 void __init
kmem_cache_init_late(void)
1631 struct kmem_cache
*cachep
;
1635 /* Annotate slab for lockdep -- annotate the malloc caches */
1638 /* 6) resize the head arrays to their final sizes */
1639 mutex_lock(&slab_mutex
);
1640 list_for_each_entry(cachep
, &slab_caches
, list
)
1641 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1643 mutex_unlock(&slab_mutex
);
1649 * Register a cpu startup notifier callback that initializes
1650 * cpu_cache_get for all new cpus
1652 register_cpu_notifier(&cpucache_notifier
);
1656 * Register a memory hotplug callback that initializes and frees
1659 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1663 * The reap timers are started later, with a module init call: That part
1664 * of the kernel is not yet operational.
1668 static int __init
cpucache_init(void)
1673 * Register the timers that return unneeded pages to the page allocator
1675 for_each_online_cpu(cpu
)
1676 start_cpu_timer(cpu
);
1682 __initcall(cpucache_init
);
1684 static noinline
void
1685 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1687 struct kmem_list3
*l3
;
1689 unsigned long flags
;
1693 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1695 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1696 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1698 for_each_online_node(node
) {
1699 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1700 unsigned long active_slabs
= 0, num_slabs
= 0;
1702 l3
= cachep
->nodelists
[node
];
1706 spin_lock_irqsave(&l3
->list_lock
, flags
);
1707 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
1708 active_objs
+= cachep
->num
;
1711 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
1712 active_objs
+= slabp
->inuse
;
1715 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
1718 free_objects
+= l3
->free_objects
;
1719 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
1721 num_slabs
+= active_slabs
;
1722 num_objs
= num_slabs
* cachep
->num
;
1724 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1725 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1731 * Interface to system's page allocator. No need to hold the cache-lock.
1733 * If we requested dmaable memory, we will get it. Even if we
1734 * did not request dmaable memory, we might get it, but that
1735 * would be relatively rare and ignorable.
1737 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1745 * Nommu uses slab's for process anonymous memory allocations, and thus
1746 * requires __GFP_COMP to properly refcount higher order allocations
1748 flags
|= __GFP_COMP
;
1751 flags
|= cachep
->allocflags
;
1752 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1753 flags
|= __GFP_RECLAIMABLE
;
1755 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1757 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1758 slab_out_of_memory(cachep
, flags
, nodeid
);
1762 nr_pages
= (1 << cachep
->gfporder
);
1763 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1764 add_zone_page_state(page_zone(page
),
1765 NR_SLAB_RECLAIMABLE
, nr_pages
);
1767 add_zone_page_state(page_zone(page
),
1768 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1769 for (i
= 0; i
< nr_pages
; i
++)
1770 __SetPageSlab(page
+ i
);
1772 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1773 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1776 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1778 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1781 return page_address(page
);
1785 * Interface to system's page release.
1787 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1789 unsigned long i
= (1 << cachep
->gfporder
);
1790 struct page
*page
= virt_to_page(addr
);
1791 const unsigned long nr_freed
= i
;
1793 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1795 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1796 sub_zone_page_state(page_zone(page
),
1797 NR_SLAB_RECLAIMABLE
, nr_freed
);
1799 sub_zone_page_state(page_zone(page
),
1800 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1802 BUG_ON(!PageSlab(page
));
1803 __ClearPageSlab(page
);
1806 if (current
->reclaim_state
)
1807 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1808 free_pages((unsigned long)addr
, cachep
->gfporder
);
1811 static void kmem_rcu_free(struct rcu_head
*head
)
1813 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1814 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1816 kmem_freepages(cachep
, slab_rcu
->addr
);
1817 if (OFF_SLAB(cachep
))
1818 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1823 #ifdef CONFIG_DEBUG_PAGEALLOC
1824 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1825 unsigned long caller
)
1827 int size
= cachep
->object_size
;
1829 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1831 if (size
< 5 * sizeof(unsigned long))
1834 *addr
++ = 0x12345678;
1836 *addr
++ = smp_processor_id();
1837 size
-= 3 * sizeof(unsigned long);
1839 unsigned long *sptr
= &caller
;
1840 unsigned long svalue
;
1842 while (!kstack_end(sptr
)) {
1844 if (kernel_text_address(svalue
)) {
1846 size
-= sizeof(unsigned long);
1847 if (size
<= sizeof(unsigned long))
1853 *addr
++ = 0x87654321;
1857 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1859 int size
= cachep
->object_size
;
1860 addr
= &((char *)addr
)[obj_offset(cachep
)];
1862 memset(addr
, val
, size
);
1863 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1866 static void dump_line(char *data
, int offset
, int limit
)
1869 unsigned char error
= 0;
1872 printk(KERN_ERR
"%03x: ", offset
);
1873 for (i
= 0; i
< limit
; i
++) {
1874 if (data
[offset
+ i
] != POISON_FREE
) {
1875 error
= data
[offset
+ i
];
1879 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1880 &data
[offset
], limit
, 1);
1882 if (bad_count
== 1) {
1883 error
^= POISON_FREE
;
1884 if (!(error
& (error
- 1))) {
1885 printk(KERN_ERR
"Single bit error detected. Probably "
1888 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1891 printk(KERN_ERR
"Run a memory test tool.\n");
1900 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1905 if (cachep
->flags
& SLAB_RED_ZONE
) {
1906 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1907 *dbg_redzone1(cachep
, objp
),
1908 *dbg_redzone2(cachep
, objp
));
1911 if (cachep
->flags
& SLAB_STORE_USER
) {
1912 printk(KERN_ERR
"Last user: [<%p>]",
1913 *dbg_userword(cachep
, objp
));
1914 print_symbol("(%s)",
1915 (unsigned long)*dbg_userword(cachep
, objp
));
1918 realobj
= (char *)objp
+ obj_offset(cachep
);
1919 size
= cachep
->object_size
;
1920 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1923 if (i
+ limit
> size
)
1925 dump_line(realobj
, i
, limit
);
1929 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1935 realobj
= (char *)objp
+ obj_offset(cachep
);
1936 size
= cachep
->object_size
;
1938 for (i
= 0; i
< size
; i
++) {
1939 char exp
= POISON_FREE
;
1942 if (realobj
[i
] != exp
) {
1948 "Slab corruption (%s): %s start=%p, len=%d\n",
1949 print_tainted(), cachep
->name
, realobj
, size
);
1950 print_objinfo(cachep
, objp
, 0);
1952 /* Hexdump the affected line */
1955 if (i
+ limit
> size
)
1957 dump_line(realobj
, i
, limit
);
1960 /* Limit to 5 lines */
1966 /* Print some data about the neighboring objects, if they
1969 struct slab
*slabp
= virt_to_slab(objp
);
1972 objnr
= obj_to_index(cachep
, slabp
, objp
);
1974 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1975 realobj
= (char *)objp
+ obj_offset(cachep
);
1976 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1978 print_objinfo(cachep
, objp
, 2);
1980 if (objnr
+ 1 < cachep
->num
) {
1981 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1982 realobj
= (char *)objp
+ obj_offset(cachep
);
1983 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1985 print_objinfo(cachep
, objp
, 2);
1992 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1995 for (i
= 0; i
< cachep
->num
; i
++) {
1996 void *objp
= index_to_obj(cachep
, slabp
, i
);
1998 if (cachep
->flags
& SLAB_POISON
) {
1999 #ifdef CONFIG_DEBUG_PAGEALLOC
2000 if (cachep
->size
% PAGE_SIZE
== 0 &&
2002 kernel_map_pages(virt_to_page(objp
),
2003 cachep
->size
/ PAGE_SIZE
, 1);
2005 check_poison_obj(cachep
, objp
);
2007 check_poison_obj(cachep
, objp
);
2010 if (cachep
->flags
& SLAB_RED_ZONE
) {
2011 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2012 slab_error(cachep
, "start of a freed object "
2014 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2015 slab_error(cachep
, "end of a freed object "
2021 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2027 * slab_destroy - destroy and release all objects in a slab
2028 * @cachep: cache pointer being destroyed
2029 * @slabp: slab pointer being destroyed
2031 * Destroy all the objs in a slab, and release the mem back to the system.
2032 * Before calling the slab must have been unlinked from the cache. The
2033 * cache-lock is not held/needed.
2035 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2037 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2039 slab_destroy_debugcheck(cachep
, slabp
);
2040 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2041 struct slab_rcu
*slab_rcu
;
2043 slab_rcu
= (struct slab_rcu
*)slabp
;
2044 slab_rcu
->cachep
= cachep
;
2045 slab_rcu
->addr
= addr
;
2046 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2048 kmem_freepages(cachep
, addr
);
2049 if (OFF_SLAB(cachep
))
2050 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2054 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
2057 struct kmem_list3
*l3
;
2059 for_each_online_cpu(i
)
2060 kfree(cachep
->array
[i
]);
2062 /* NUMA: free the list3 structures */
2063 for_each_online_node(i
) {
2064 l3
= cachep
->nodelists
[i
];
2067 free_alien_cache(l3
->alien
);
2071 kmem_cache_free(&cache_cache
, cachep
);
2076 * calculate_slab_order - calculate size (page order) of slabs
2077 * @cachep: pointer to the cache that is being created
2078 * @size: size of objects to be created in this cache.
2079 * @align: required alignment for the objects.
2080 * @flags: slab allocation flags
2082 * Also calculates the number of objects per slab.
2084 * This could be made much more intelligent. For now, try to avoid using
2085 * high order pages for slabs. When the gfp() functions are more friendly
2086 * towards high-order requests, this should be changed.
2088 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2089 size_t size
, size_t align
, unsigned long flags
)
2091 unsigned long offslab_limit
;
2092 size_t left_over
= 0;
2095 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2099 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2103 if (flags
& CFLGS_OFF_SLAB
) {
2105 * Max number of objs-per-slab for caches which
2106 * use off-slab slabs. Needed to avoid a possible
2107 * looping condition in cache_grow().
2109 offslab_limit
= size
- sizeof(struct slab
);
2110 offslab_limit
/= sizeof(kmem_bufctl_t
);
2112 if (num
> offslab_limit
)
2116 /* Found something acceptable - save it away */
2118 cachep
->gfporder
= gfporder
;
2119 left_over
= remainder
;
2122 * A VFS-reclaimable slab tends to have most allocations
2123 * as GFP_NOFS and we really don't want to have to be allocating
2124 * higher-order pages when we are unable to shrink dcache.
2126 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2130 * Large number of objects is good, but very large slabs are
2131 * currently bad for the gfp()s.
2133 if (gfporder
>= slab_max_order
)
2137 * Acceptable internal fragmentation?
2139 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2145 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2147 if (slab_state
>= FULL
)
2148 return enable_cpucache(cachep
, gfp
);
2150 if (slab_state
== DOWN
) {
2152 * Note: the first kmem_cache_create must create the cache
2153 * that's used by kmalloc(24), otherwise the creation of
2154 * further caches will BUG().
2156 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2159 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2160 * the first cache, then we need to set up all its list3s,
2161 * otherwise the creation of further caches will BUG().
2163 set_up_list3s(cachep
, SIZE_AC
);
2164 if (INDEX_AC
== INDEX_L3
)
2165 slab_state
= PARTIAL_L3
;
2167 slab_state
= PARTIAL_ARRAYCACHE
;
2169 cachep
->array
[smp_processor_id()] =
2170 kmalloc(sizeof(struct arraycache_init
), gfp
);
2172 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2173 set_up_list3s(cachep
, SIZE_L3
);
2174 slab_state
= PARTIAL_L3
;
2177 for_each_online_node(node
) {
2178 cachep
->nodelists
[node
] =
2179 kmalloc_node(sizeof(struct kmem_list3
),
2181 BUG_ON(!cachep
->nodelists
[node
]);
2182 kmem_list3_init(cachep
->nodelists
[node
]);
2186 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2187 jiffies
+ REAPTIMEOUT_LIST3
+
2188 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2190 cpu_cache_get(cachep
)->avail
= 0;
2191 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2192 cpu_cache_get(cachep
)->batchcount
= 1;
2193 cpu_cache_get(cachep
)->touched
= 0;
2194 cachep
->batchcount
= 1;
2195 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2200 * __kmem_cache_create - Create a cache.
2201 * @name: A string which is used in /proc/slabinfo to identify this cache.
2202 * @size: The size of objects to be created in this cache.
2203 * @align: The required alignment for the objects.
2204 * @flags: SLAB flags
2205 * @ctor: A constructor for the objects.
2207 * Returns a ptr to the cache on success, NULL on failure.
2208 * Cannot be called within a int, but can be interrupted.
2209 * The @ctor is run when new pages are allocated by the cache.
2211 * @name must be valid until the cache is destroyed. This implies that
2212 * the module calling this has to destroy the cache before getting unloaded.
2216 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2217 * to catch references to uninitialised memory.
2219 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2220 * for buffer overruns.
2222 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2223 * cacheline. This can be beneficial if you're counting cycles as closely
2227 __kmem_cache_create (const char *name
, size_t size
, size_t align
,
2228 unsigned long flags
, void (*ctor
)(void *))
2230 size_t left_over
, slab_size
, ralign
;
2231 struct kmem_cache
*cachep
= NULL
;
2237 * Enable redzoning and last user accounting, except for caches with
2238 * large objects, if the increased size would increase the object size
2239 * above the next power of two: caches with object sizes just above a
2240 * power of two have a significant amount of internal fragmentation.
2242 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2243 2 * sizeof(unsigned long long)))
2244 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2245 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2246 flags
|= SLAB_POISON
;
2248 if (flags
& SLAB_DESTROY_BY_RCU
)
2249 BUG_ON(flags
& SLAB_POISON
);
2252 * Always checks flags, a caller might be expecting debug support which
2255 BUG_ON(flags
& ~CREATE_MASK
);
2258 * Check that size is in terms of words. This is needed to avoid
2259 * unaligned accesses for some archs when redzoning is used, and makes
2260 * sure any on-slab bufctl's are also correctly aligned.
2262 if (size
& (BYTES_PER_WORD
- 1)) {
2263 size
+= (BYTES_PER_WORD
- 1);
2264 size
&= ~(BYTES_PER_WORD
- 1);
2267 /* calculate the final buffer alignment: */
2269 /* 1) arch recommendation: can be overridden for debug */
2270 if (flags
& SLAB_HWCACHE_ALIGN
) {
2272 * Default alignment: as specified by the arch code. Except if
2273 * an object is really small, then squeeze multiple objects into
2276 ralign
= cache_line_size();
2277 while (size
<= ralign
/ 2)
2280 ralign
= BYTES_PER_WORD
;
2284 * Redzoning and user store require word alignment or possibly larger.
2285 * Note this will be overridden by architecture or caller mandated
2286 * alignment if either is greater than BYTES_PER_WORD.
2288 if (flags
& SLAB_STORE_USER
)
2289 ralign
= BYTES_PER_WORD
;
2291 if (flags
& SLAB_RED_ZONE
) {
2292 ralign
= REDZONE_ALIGN
;
2293 /* If redzoning, ensure that the second redzone is suitably
2294 * aligned, by adjusting the object size accordingly. */
2295 size
+= REDZONE_ALIGN
- 1;
2296 size
&= ~(REDZONE_ALIGN
- 1);
2299 /* 2) arch mandated alignment */
2300 if (ralign
< ARCH_SLAB_MINALIGN
) {
2301 ralign
= ARCH_SLAB_MINALIGN
;
2303 /* 3) caller mandated alignment */
2304 if (ralign
< align
) {
2307 /* disable debug if necessary */
2308 if (ralign
> __alignof__(unsigned long long))
2309 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2315 if (slab_is_available())
2320 /* Get cache's description obj. */
2321 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2325 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
2326 cachep
->object_size
= size
;
2327 cachep
->align
= align
;
2331 * Both debugging options require word-alignment which is calculated
2334 if (flags
& SLAB_RED_ZONE
) {
2335 /* add space for red zone words */
2336 cachep
->obj_offset
+= sizeof(unsigned long long);
2337 size
+= 2 * sizeof(unsigned long long);
2339 if (flags
& SLAB_STORE_USER
) {
2340 /* user store requires one word storage behind the end of
2341 * the real object. But if the second red zone needs to be
2342 * aligned to 64 bits, we must allow that much space.
2344 if (flags
& SLAB_RED_ZONE
)
2345 size
+= REDZONE_ALIGN
;
2347 size
+= BYTES_PER_WORD
;
2349 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2350 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2351 && cachep
->object_size
> cache_line_size() && ALIGN(size
, align
) < PAGE_SIZE
) {
2352 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, align
);
2359 * Determine if the slab management is 'on' or 'off' slab.
2360 * (bootstrapping cannot cope with offslab caches so don't do
2361 * it too early on. Always use on-slab management when
2362 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2364 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2365 !(flags
& SLAB_NOLEAKTRACE
))
2367 * Size is large, assume best to place the slab management obj
2368 * off-slab (should allow better packing of objs).
2370 flags
|= CFLGS_OFF_SLAB
;
2372 size
= ALIGN(size
, align
);
2374 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2378 "kmem_cache_create: couldn't create cache %s.\n", name
);
2379 kmem_cache_free(&cache_cache
, cachep
);
2382 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2383 + sizeof(struct slab
), align
);
2386 * If the slab has been placed off-slab, and we have enough space then
2387 * move it on-slab. This is at the expense of any extra colouring.
2389 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2390 flags
&= ~CFLGS_OFF_SLAB
;
2391 left_over
-= slab_size
;
2394 if (flags
& CFLGS_OFF_SLAB
) {
2395 /* really off slab. No need for manual alignment */
2397 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2399 #ifdef CONFIG_PAGE_POISONING
2400 /* If we're going to use the generic kernel_map_pages()
2401 * poisoning, then it's going to smash the contents of
2402 * the redzone and userword anyhow, so switch them off.
2404 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2405 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2409 cachep
->colour_off
= cache_line_size();
2410 /* Offset must be a multiple of the alignment. */
2411 if (cachep
->colour_off
< align
)
2412 cachep
->colour_off
= align
;
2413 cachep
->colour
= left_over
/ cachep
->colour_off
;
2414 cachep
->slab_size
= slab_size
;
2415 cachep
->flags
= flags
;
2416 cachep
->allocflags
= 0;
2417 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2418 cachep
->allocflags
|= GFP_DMA
;
2419 cachep
->size
= size
;
2420 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2422 if (flags
& CFLGS_OFF_SLAB
) {
2423 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2425 * This is a possibility for one of the malloc_sizes caches.
2426 * But since we go off slab only for object size greater than
2427 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2428 * this should not happen at all.
2429 * But leave a BUG_ON for some lucky dude.
2431 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2433 cachep
->ctor
= ctor
;
2434 cachep
->name
= name
;
2436 if (setup_cpu_cache(cachep
, gfp
)) {
2437 __kmem_cache_destroy(cachep
);
2441 if (flags
& SLAB_DEBUG_OBJECTS
) {
2443 * Would deadlock through slab_destroy()->call_rcu()->
2444 * debug_object_activate()->kmem_cache_alloc().
2446 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2448 slab_set_debugobj_lock_classes(cachep
);
2451 /* cache setup completed, link it into the list */
2452 list_add(&cachep
->list
, &slab_caches
);
2457 static void check_irq_off(void)
2459 BUG_ON(!irqs_disabled());
2462 static void check_irq_on(void)
2464 BUG_ON(irqs_disabled());
2467 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2471 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2475 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2479 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2484 #define check_irq_off() do { } while(0)
2485 #define check_irq_on() do { } while(0)
2486 #define check_spinlock_acquired(x) do { } while(0)
2487 #define check_spinlock_acquired_node(x, y) do { } while(0)
2490 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2491 struct array_cache
*ac
,
2492 int force
, int node
);
2494 static void do_drain(void *arg
)
2496 struct kmem_cache
*cachep
= arg
;
2497 struct array_cache
*ac
;
2498 int node
= numa_mem_id();
2501 ac
= cpu_cache_get(cachep
);
2502 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2503 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2504 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2508 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2510 struct kmem_list3
*l3
;
2513 on_each_cpu(do_drain
, cachep
, 1);
2515 for_each_online_node(node
) {
2516 l3
= cachep
->nodelists
[node
];
2517 if (l3
&& l3
->alien
)
2518 drain_alien_cache(cachep
, l3
->alien
);
2521 for_each_online_node(node
) {
2522 l3
= cachep
->nodelists
[node
];
2524 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2529 * Remove slabs from the list of free slabs.
2530 * Specify the number of slabs to drain in tofree.
2532 * Returns the actual number of slabs released.
2534 static int drain_freelist(struct kmem_cache
*cache
,
2535 struct kmem_list3
*l3
, int tofree
)
2537 struct list_head
*p
;
2542 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2544 spin_lock_irq(&l3
->list_lock
);
2545 p
= l3
->slabs_free
.prev
;
2546 if (p
== &l3
->slabs_free
) {
2547 spin_unlock_irq(&l3
->list_lock
);
2551 slabp
= list_entry(p
, struct slab
, list
);
2553 BUG_ON(slabp
->inuse
);
2555 list_del(&slabp
->list
);
2557 * Safe to drop the lock. The slab is no longer linked
2560 l3
->free_objects
-= cache
->num
;
2561 spin_unlock_irq(&l3
->list_lock
);
2562 slab_destroy(cache
, slabp
);
2569 /* Called with slab_mutex held to protect against cpu hotplug */
2570 static int __cache_shrink(struct kmem_cache
*cachep
)
2573 struct kmem_list3
*l3
;
2575 drain_cpu_caches(cachep
);
2578 for_each_online_node(i
) {
2579 l3
= cachep
->nodelists
[i
];
2583 drain_freelist(cachep
, l3
, l3
->free_objects
);
2585 ret
+= !list_empty(&l3
->slabs_full
) ||
2586 !list_empty(&l3
->slabs_partial
);
2588 return (ret
? 1 : 0);
2592 * kmem_cache_shrink - Shrink a cache.
2593 * @cachep: The cache to shrink.
2595 * Releases as many slabs as possible for a cache.
2596 * To help debugging, a zero exit status indicates all slabs were released.
2598 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2601 BUG_ON(!cachep
|| in_interrupt());
2604 mutex_lock(&slab_mutex
);
2605 ret
= __cache_shrink(cachep
);
2606 mutex_unlock(&slab_mutex
);
2610 EXPORT_SYMBOL(kmem_cache_shrink
);
2613 * kmem_cache_destroy - delete a cache
2614 * @cachep: the cache to destroy
2616 * Remove a &struct kmem_cache object from the slab cache.
2618 * It is expected this function will be called by a module when it is
2619 * unloaded. This will remove the cache completely, and avoid a duplicate
2620 * cache being allocated each time a module is loaded and unloaded, if the
2621 * module doesn't have persistent in-kernel storage across loads and unloads.
2623 * The cache must be empty before calling this function.
2625 * The caller must guarantee that no one will allocate memory from the cache
2626 * during the kmem_cache_destroy().
2628 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2630 BUG_ON(!cachep
|| in_interrupt());
2632 /* Find the cache in the chain of caches. */
2634 mutex_lock(&slab_mutex
);
2636 * the chain is never empty, cache_cache is never destroyed
2638 list_del(&cachep
->list
);
2639 if (__cache_shrink(cachep
)) {
2640 slab_error(cachep
, "Can't free all objects");
2641 list_add(&cachep
->list
, &slab_caches
);
2642 mutex_unlock(&slab_mutex
);
2647 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2650 __kmem_cache_destroy(cachep
);
2651 mutex_unlock(&slab_mutex
);
2654 EXPORT_SYMBOL(kmem_cache_destroy
);
2657 * Get the memory for a slab management obj.
2658 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2659 * always come from malloc_sizes caches. The slab descriptor cannot
2660 * come from the same cache which is getting created because,
2661 * when we are searching for an appropriate cache for these
2662 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2663 * If we are creating a malloc_sizes cache here it would not be visible to
2664 * kmem_find_general_cachep till the initialization is complete.
2665 * Hence we cannot have slabp_cache same as the original cache.
2667 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2668 int colour_off
, gfp_t local_flags
,
2673 if (OFF_SLAB(cachep
)) {
2674 /* Slab management obj is off-slab. */
2675 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2676 local_flags
, nodeid
);
2678 * If the first object in the slab is leaked (it's allocated
2679 * but no one has a reference to it), we want to make sure
2680 * kmemleak does not treat the ->s_mem pointer as a reference
2681 * to the object. Otherwise we will not report the leak.
2683 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2688 slabp
= objp
+ colour_off
;
2689 colour_off
+= cachep
->slab_size
;
2692 slabp
->colouroff
= colour_off
;
2693 slabp
->s_mem
= objp
+ colour_off
;
2694 slabp
->nodeid
= nodeid
;
2699 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2701 return (kmem_bufctl_t
*) (slabp
+ 1);
2704 static void cache_init_objs(struct kmem_cache
*cachep
,
2709 for (i
= 0; i
< cachep
->num
; i
++) {
2710 void *objp
= index_to_obj(cachep
, slabp
, i
);
2712 /* need to poison the objs? */
2713 if (cachep
->flags
& SLAB_POISON
)
2714 poison_obj(cachep
, objp
, POISON_FREE
);
2715 if (cachep
->flags
& SLAB_STORE_USER
)
2716 *dbg_userword(cachep
, objp
) = NULL
;
2718 if (cachep
->flags
& SLAB_RED_ZONE
) {
2719 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2720 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2723 * Constructors are not allowed to allocate memory from the same
2724 * cache which they are a constructor for. Otherwise, deadlock.
2725 * They must also be threaded.
2727 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2728 cachep
->ctor(objp
+ obj_offset(cachep
));
2730 if (cachep
->flags
& SLAB_RED_ZONE
) {
2731 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2732 slab_error(cachep
, "constructor overwrote the"
2733 " end of an object");
2734 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2735 slab_error(cachep
, "constructor overwrote the"
2736 " start of an object");
2738 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2739 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2740 kernel_map_pages(virt_to_page(objp
),
2741 cachep
->size
/ PAGE_SIZE
, 0);
2746 slab_bufctl(slabp
)[i
] = i
+ 1;
2748 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2751 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2753 if (CONFIG_ZONE_DMA_FLAG
) {
2754 if (flags
& GFP_DMA
)
2755 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2757 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2761 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2764 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2768 next
= slab_bufctl(slabp
)[slabp
->free
];
2770 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2771 WARN_ON(slabp
->nodeid
!= nodeid
);
2778 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2779 void *objp
, int nodeid
)
2781 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2784 /* Verify that the slab belongs to the intended node */
2785 WARN_ON(slabp
->nodeid
!= nodeid
);
2787 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2788 printk(KERN_ERR
"slab: double free detected in cache "
2789 "'%s', objp %p\n", cachep
->name
, objp
);
2793 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2794 slabp
->free
= objnr
;
2799 * Map pages beginning at addr to the given cache and slab. This is required
2800 * for the slab allocator to be able to lookup the cache and slab of a
2801 * virtual address for kfree, ksize, and slab debugging.
2803 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2809 page
= virt_to_page(addr
);
2812 if (likely(!PageCompound(page
)))
2813 nr_pages
<<= cache
->gfporder
;
2816 page
->slab_cache
= cache
;
2817 page
->slab_page
= slab
;
2819 } while (--nr_pages
);
2823 * Grow (by 1) the number of slabs within a cache. This is called by
2824 * kmem_cache_alloc() when there are no active objs left in a cache.
2826 static int cache_grow(struct kmem_cache
*cachep
,
2827 gfp_t flags
, int nodeid
, void *objp
)
2832 struct kmem_list3
*l3
;
2835 * Be lazy and only check for valid flags here, keeping it out of the
2836 * critical path in kmem_cache_alloc().
2838 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2839 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2841 /* Take the l3 list lock to change the colour_next on this node */
2843 l3
= cachep
->nodelists
[nodeid
];
2844 spin_lock(&l3
->list_lock
);
2846 /* Get colour for the slab, and cal the next value. */
2847 offset
= l3
->colour_next
;
2849 if (l3
->colour_next
>= cachep
->colour
)
2850 l3
->colour_next
= 0;
2851 spin_unlock(&l3
->list_lock
);
2853 offset
*= cachep
->colour_off
;
2855 if (local_flags
& __GFP_WAIT
)
2859 * The test for missing atomic flag is performed here, rather than
2860 * the more obvious place, simply to reduce the critical path length
2861 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2862 * will eventually be caught here (where it matters).
2864 kmem_flagcheck(cachep
, flags
);
2867 * Get mem for the objs. Attempt to allocate a physical page from
2871 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2875 /* Get slab management. */
2876 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2877 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2881 slab_map_pages(cachep
, slabp
, objp
);
2883 cache_init_objs(cachep
, slabp
);
2885 if (local_flags
& __GFP_WAIT
)
2886 local_irq_disable();
2888 spin_lock(&l3
->list_lock
);
2890 /* Make slab active. */
2891 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2892 STATS_INC_GROWN(cachep
);
2893 l3
->free_objects
+= cachep
->num
;
2894 spin_unlock(&l3
->list_lock
);
2897 kmem_freepages(cachep
, objp
);
2899 if (local_flags
& __GFP_WAIT
)
2900 local_irq_disable();
2907 * Perform extra freeing checks:
2908 * - detect bad pointers.
2909 * - POISON/RED_ZONE checking
2911 static void kfree_debugcheck(const void *objp
)
2913 if (!virt_addr_valid(objp
)) {
2914 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2915 (unsigned long)objp
);
2920 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2922 unsigned long long redzone1
, redzone2
;
2924 redzone1
= *dbg_redzone1(cache
, obj
);
2925 redzone2
= *dbg_redzone2(cache
, obj
);
2930 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2933 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2934 slab_error(cache
, "double free detected");
2936 slab_error(cache
, "memory outside object was overwritten");
2938 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2939 obj
, redzone1
, redzone2
);
2942 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2949 BUG_ON(virt_to_cache(objp
) != cachep
);
2951 objp
-= obj_offset(cachep
);
2952 kfree_debugcheck(objp
);
2953 page
= virt_to_head_page(objp
);
2955 slabp
= page
->slab_page
;
2957 if (cachep
->flags
& SLAB_RED_ZONE
) {
2958 verify_redzone_free(cachep
, objp
);
2959 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2960 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2962 if (cachep
->flags
& SLAB_STORE_USER
)
2963 *dbg_userword(cachep
, objp
) = caller
;
2965 objnr
= obj_to_index(cachep
, slabp
, objp
);
2967 BUG_ON(objnr
>= cachep
->num
);
2968 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2970 #ifdef CONFIG_DEBUG_SLAB_LEAK
2971 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2973 if (cachep
->flags
& SLAB_POISON
) {
2974 #ifdef CONFIG_DEBUG_PAGEALLOC
2975 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2976 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2977 kernel_map_pages(virt_to_page(objp
),
2978 cachep
->size
/ PAGE_SIZE
, 0);
2980 poison_obj(cachep
, objp
, POISON_FREE
);
2983 poison_obj(cachep
, objp
, POISON_FREE
);
2989 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2994 /* Check slab's freelist to see if this obj is there. */
2995 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2997 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3000 if (entries
!= cachep
->num
- slabp
->inuse
) {
3002 printk(KERN_ERR
"slab: Internal list corruption detected in "
3003 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3004 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
3006 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3007 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3013 #define kfree_debugcheck(x) do { } while(0)
3014 #define cache_free_debugcheck(x,objp,z) (objp)
3015 #define check_slabp(x,y) do { } while(0)
3018 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
3021 struct kmem_list3
*l3
;
3022 struct array_cache
*ac
;
3027 node
= numa_mem_id();
3028 ac
= cpu_cache_get(cachep
);
3029 batchcount
= ac
->batchcount
;
3030 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3032 * If there was little recent activity on this cache, then
3033 * perform only a partial refill. Otherwise we could generate
3036 batchcount
= BATCHREFILL_LIMIT
;
3038 l3
= cachep
->nodelists
[node
];
3040 BUG_ON(ac
->avail
> 0 || !l3
);
3041 spin_lock(&l3
->list_lock
);
3043 /* See if we can refill from the shared array */
3044 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3045 l3
->shared
->touched
= 1;
3049 while (batchcount
> 0) {
3050 struct list_head
*entry
;
3052 /* Get slab alloc is to come from. */
3053 entry
= l3
->slabs_partial
.next
;
3054 if (entry
== &l3
->slabs_partial
) {
3055 l3
->free_touched
= 1;
3056 entry
= l3
->slabs_free
.next
;
3057 if (entry
== &l3
->slabs_free
)
3061 slabp
= list_entry(entry
, struct slab
, list
);
3062 check_slabp(cachep
, slabp
);
3063 check_spinlock_acquired(cachep
);
3066 * The slab was either on partial or free list so
3067 * there must be at least one object available for
3070 BUG_ON(slabp
->inuse
>= cachep
->num
);
3072 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3073 STATS_INC_ALLOCED(cachep
);
3074 STATS_INC_ACTIVE(cachep
);
3075 STATS_SET_HIGH(cachep
);
3077 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
3080 check_slabp(cachep
, slabp
);
3082 /* move slabp to correct slabp list: */
3083 list_del(&slabp
->list
);
3084 if (slabp
->free
== BUFCTL_END
)
3085 list_add(&slabp
->list
, &l3
->slabs_full
);
3087 list_add(&slabp
->list
, &l3
->slabs_partial
);
3091 l3
->free_objects
-= ac
->avail
;
3093 spin_unlock(&l3
->list_lock
);
3095 if (unlikely(!ac
->avail
)) {
3097 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3099 /* cache_grow can reenable interrupts, then ac could change. */
3100 ac
= cpu_cache_get(cachep
);
3101 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3104 if (!ac
->avail
) /* objects refilled by interrupt? */
3108 return ac
->entry
[--ac
->avail
];
3111 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3114 might_sleep_if(flags
& __GFP_WAIT
);
3116 kmem_flagcheck(cachep
, flags
);
3121 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3122 gfp_t flags
, void *objp
, void *caller
)
3126 if (cachep
->flags
& SLAB_POISON
) {
3127 #ifdef CONFIG_DEBUG_PAGEALLOC
3128 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3129 kernel_map_pages(virt_to_page(objp
),
3130 cachep
->size
/ PAGE_SIZE
, 1);
3132 check_poison_obj(cachep
, objp
);
3134 check_poison_obj(cachep
, objp
);
3136 poison_obj(cachep
, objp
, POISON_INUSE
);
3138 if (cachep
->flags
& SLAB_STORE_USER
)
3139 *dbg_userword(cachep
, objp
) = caller
;
3141 if (cachep
->flags
& SLAB_RED_ZONE
) {
3142 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3143 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3144 slab_error(cachep
, "double free, or memory outside"
3145 " object was overwritten");
3147 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3148 objp
, *dbg_redzone1(cachep
, objp
),
3149 *dbg_redzone2(cachep
, objp
));
3151 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3152 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3154 #ifdef CONFIG_DEBUG_SLAB_LEAK
3159 slabp
= virt_to_head_page(objp
)->slab_page
;
3160 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3161 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3164 objp
+= obj_offset(cachep
);
3165 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3167 if (ARCH_SLAB_MINALIGN
&&
3168 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3169 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3170 objp
, (int)ARCH_SLAB_MINALIGN
);
3175 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3178 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3180 if (cachep
== &cache_cache
)
3183 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3186 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3189 struct array_cache
*ac
;
3193 ac
= cpu_cache_get(cachep
);
3194 if (likely(ac
->avail
)) {
3195 STATS_INC_ALLOCHIT(cachep
);
3197 objp
= ac
->entry
[--ac
->avail
];
3199 STATS_INC_ALLOCMISS(cachep
);
3200 objp
= cache_alloc_refill(cachep
, flags
);
3202 * the 'ac' may be updated by cache_alloc_refill(),
3203 * and kmemleak_erase() requires its correct value.
3205 ac
= cpu_cache_get(cachep
);
3208 * To avoid a false negative, if an object that is in one of the
3209 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3210 * treat the array pointers as a reference to the object.
3213 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3219 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3221 * If we are in_interrupt, then process context, including cpusets and
3222 * mempolicy, may not apply and should not be used for allocation policy.
3224 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3226 int nid_alloc
, nid_here
;
3228 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3230 nid_alloc
= nid_here
= numa_mem_id();
3231 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3232 nid_alloc
= cpuset_slab_spread_node();
3233 else if (current
->mempolicy
)
3234 nid_alloc
= slab_node();
3235 if (nid_alloc
!= nid_here
)
3236 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3241 * Fallback function if there was no memory available and no objects on a
3242 * certain node and fall back is permitted. First we scan all the
3243 * available nodelists for available objects. If that fails then we
3244 * perform an allocation without specifying a node. This allows the page
3245 * allocator to do its reclaim / fallback magic. We then insert the
3246 * slab into the proper nodelist and then allocate from it.
3248 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3250 struct zonelist
*zonelist
;
3254 enum zone_type high_zoneidx
= gfp_zone(flags
);
3257 unsigned int cpuset_mems_cookie
;
3259 if (flags
& __GFP_THISNODE
)
3262 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3265 cpuset_mems_cookie
= get_mems_allowed();
3266 zonelist
= node_zonelist(slab_node(), flags
);
3270 * Look through allowed nodes for objects available
3271 * from existing per node queues.
3273 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3274 nid
= zone_to_nid(zone
);
3276 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3277 cache
->nodelists
[nid
] &&
3278 cache
->nodelists
[nid
]->free_objects
) {
3279 obj
= ____cache_alloc_node(cache
,
3280 flags
| GFP_THISNODE
, nid
);
3288 * This allocation will be performed within the constraints
3289 * of the current cpuset / memory policy requirements.
3290 * We may trigger various forms of reclaim on the allowed
3291 * set and go into memory reserves if necessary.
3293 if (local_flags
& __GFP_WAIT
)
3295 kmem_flagcheck(cache
, flags
);
3296 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3297 if (local_flags
& __GFP_WAIT
)
3298 local_irq_disable();
3301 * Insert into the appropriate per node queues
3303 nid
= page_to_nid(virt_to_page(obj
));
3304 if (cache_grow(cache
, flags
, nid
, obj
)) {
3305 obj
= ____cache_alloc_node(cache
,
3306 flags
| GFP_THISNODE
, nid
);
3309 * Another processor may allocate the
3310 * objects in the slab since we are
3311 * not holding any locks.
3315 /* cache_grow already freed obj */
3321 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3327 * A interface to enable slab creation on nodeid
3329 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3332 struct list_head
*entry
;
3334 struct kmem_list3
*l3
;
3338 l3
= cachep
->nodelists
[nodeid
];
3343 spin_lock(&l3
->list_lock
);
3344 entry
= l3
->slabs_partial
.next
;
3345 if (entry
== &l3
->slabs_partial
) {
3346 l3
->free_touched
= 1;
3347 entry
= l3
->slabs_free
.next
;
3348 if (entry
== &l3
->slabs_free
)
3352 slabp
= list_entry(entry
, struct slab
, list
);
3353 check_spinlock_acquired_node(cachep
, nodeid
);
3354 check_slabp(cachep
, slabp
);
3356 STATS_INC_NODEALLOCS(cachep
);
3357 STATS_INC_ACTIVE(cachep
);
3358 STATS_SET_HIGH(cachep
);
3360 BUG_ON(slabp
->inuse
== cachep
->num
);
3362 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3363 check_slabp(cachep
, slabp
);
3365 /* move slabp to correct slabp list: */
3366 list_del(&slabp
->list
);
3368 if (slabp
->free
== BUFCTL_END
)
3369 list_add(&slabp
->list
, &l3
->slabs_full
);
3371 list_add(&slabp
->list
, &l3
->slabs_partial
);
3373 spin_unlock(&l3
->list_lock
);
3377 spin_unlock(&l3
->list_lock
);
3378 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3382 return fallback_alloc(cachep
, flags
);
3389 * kmem_cache_alloc_node - Allocate an object on the specified node
3390 * @cachep: The cache to allocate from.
3391 * @flags: See kmalloc().
3392 * @nodeid: node number of the target node.
3393 * @caller: return address of caller, used for debug information
3395 * Identical to kmem_cache_alloc but it will allocate memory on the given
3396 * node, which can improve the performance for cpu bound structures.
3398 * Fallback to other node is possible if __GFP_THISNODE is not set.
3400 static __always_inline
void *
3401 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3404 unsigned long save_flags
;
3406 int slab_node
= numa_mem_id();
3408 flags
&= gfp_allowed_mask
;
3410 lockdep_trace_alloc(flags
);
3412 if (slab_should_failslab(cachep
, flags
))
3415 cache_alloc_debugcheck_before(cachep
, flags
);
3416 local_irq_save(save_flags
);
3418 if (nodeid
== NUMA_NO_NODE
)
3421 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3422 /* Node not bootstrapped yet */
3423 ptr
= fallback_alloc(cachep
, flags
);
3427 if (nodeid
== slab_node
) {
3429 * Use the locally cached objects if possible.
3430 * However ____cache_alloc does not allow fallback
3431 * to other nodes. It may fail while we still have
3432 * objects on other nodes available.
3434 ptr
= ____cache_alloc(cachep
, flags
);
3438 /* ___cache_alloc_node can fall back to other nodes */
3439 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3441 local_irq_restore(save_flags
);
3442 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3443 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3447 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3449 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3450 memset(ptr
, 0, cachep
->object_size
);
3455 static __always_inline
void *
3456 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3460 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3461 objp
= alternate_node_alloc(cache
, flags
);
3465 objp
= ____cache_alloc(cache
, flags
);
3468 * We may just have run out of memory on the local node.
3469 * ____cache_alloc_node() knows how to locate memory on other nodes
3472 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3479 static __always_inline
void *
3480 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3482 return ____cache_alloc(cachep
, flags
);
3485 #endif /* CONFIG_NUMA */
3487 static __always_inline
void *
3488 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3490 unsigned long save_flags
;
3493 flags
&= gfp_allowed_mask
;
3495 lockdep_trace_alloc(flags
);
3497 if (slab_should_failslab(cachep
, flags
))
3500 cache_alloc_debugcheck_before(cachep
, flags
);
3501 local_irq_save(save_flags
);
3502 objp
= __do_cache_alloc(cachep
, flags
);
3503 local_irq_restore(save_flags
);
3504 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3505 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3510 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3512 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3513 memset(objp
, 0, cachep
->object_size
);
3519 * Caller needs to acquire correct kmem_list's list_lock
3521 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3525 struct kmem_list3
*l3
;
3527 for (i
= 0; i
< nr_objects
; i
++) {
3528 void *objp
= objpp
[i
];
3531 slabp
= virt_to_slab(objp
);
3532 l3
= cachep
->nodelists
[node
];
3533 list_del(&slabp
->list
);
3534 check_spinlock_acquired_node(cachep
, node
);
3535 check_slabp(cachep
, slabp
);
3536 slab_put_obj(cachep
, slabp
, objp
, node
);
3537 STATS_DEC_ACTIVE(cachep
);
3539 check_slabp(cachep
, slabp
);
3541 /* fixup slab chains */
3542 if (slabp
->inuse
== 0) {
3543 if (l3
->free_objects
> l3
->free_limit
) {
3544 l3
->free_objects
-= cachep
->num
;
3545 /* No need to drop any previously held
3546 * lock here, even if we have a off-slab slab
3547 * descriptor it is guaranteed to come from
3548 * a different cache, refer to comments before
3551 slab_destroy(cachep
, slabp
);
3553 list_add(&slabp
->list
, &l3
->slabs_free
);
3556 /* Unconditionally move a slab to the end of the
3557 * partial list on free - maximum time for the
3558 * other objects to be freed, too.
3560 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3565 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3568 struct kmem_list3
*l3
;
3569 int node
= numa_mem_id();
3571 batchcount
= ac
->batchcount
;
3573 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3576 l3
= cachep
->nodelists
[node
];
3577 spin_lock(&l3
->list_lock
);
3579 struct array_cache
*shared_array
= l3
->shared
;
3580 int max
= shared_array
->limit
- shared_array
->avail
;
3582 if (batchcount
> max
)
3584 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3585 ac
->entry
, sizeof(void *) * batchcount
);
3586 shared_array
->avail
+= batchcount
;
3591 free_block(cachep
, ac
->entry
, batchcount
, node
);
3596 struct list_head
*p
;
3598 p
= l3
->slabs_free
.next
;
3599 while (p
!= &(l3
->slabs_free
)) {
3602 slabp
= list_entry(p
, struct slab
, list
);
3603 BUG_ON(slabp
->inuse
);
3608 STATS_SET_FREEABLE(cachep
, i
);
3611 spin_unlock(&l3
->list_lock
);
3612 ac
->avail
-= batchcount
;
3613 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3617 * Release an obj back to its cache. If the obj has a constructed state, it must
3618 * be in this state _before_ it is released. Called with disabled ints.
3620 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3623 struct array_cache
*ac
= cpu_cache_get(cachep
);
3626 kmemleak_free_recursive(objp
, cachep
->flags
);
3627 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3629 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3632 * Skip calling cache_free_alien() when the platform is not numa.
3633 * This will avoid cache misses that happen while accessing slabp (which
3634 * is per page memory reference) to get nodeid. Instead use a global
3635 * variable to skip the call, which is mostly likely to be present in
3638 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3641 if (likely(ac
->avail
< ac
->limit
)) {
3642 STATS_INC_FREEHIT(cachep
);
3644 STATS_INC_FREEMISS(cachep
);
3645 cache_flusharray(cachep
, ac
);
3648 ac
->entry
[ac
->avail
++] = objp
;
3652 * kmem_cache_alloc - Allocate an object
3653 * @cachep: The cache to allocate from.
3654 * @flags: See kmalloc().
3656 * Allocate an object from this cache. The flags are only relevant
3657 * if the cache has no available objects.
3659 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3661 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3663 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3664 cachep
->object_size
, cachep
->size
, flags
);
3668 EXPORT_SYMBOL(kmem_cache_alloc
);
3670 #ifdef CONFIG_TRACING
3672 kmem_cache_alloc_trace(size_t size
, struct kmem_cache
*cachep
, gfp_t flags
)
3676 ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3678 trace_kmalloc(_RET_IP_
, ret
,
3679 size
, slab_buffer_size(cachep
), flags
);
3682 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3686 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3688 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3689 __builtin_return_address(0));
3691 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3692 cachep
->object_size
, cachep
->size
,
3697 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3699 #ifdef CONFIG_TRACING
3700 void *kmem_cache_alloc_node_trace(size_t size
,
3701 struct kmem_cache
*cachep
,
3707 ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3708 __builtin_return_address(0));
3709 trace_kmalloc_node(_RET_IP_
, ret
,
3710 size
, slab_buffer_size(cachep
),
3714 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3717 static __always_inline
void *
3718 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3720 struct kmem_cache
*cachep
;
3722 cachep
= kmem_find_general_cachep(size
, flags
);
3723 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3725 return kmem_cache_alloc_node_trace(size
, cachep
, flags
, node
);
3728 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3729 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3731 return __do_kmalloc_node(size
, flags
, node
,
3732 __builtin_return_address(0));
3734 EXPORT_SYMBOL(__kmalloc_node
);
3736 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3737 int node
, unsigned long caller
)
3739 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3741 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3743 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3745 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3747 EXPORT_SYMBOL(__kmalloc_node
);
3748 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3749 #endif /* CONFIG_NUMA */
3752 * __do_kmalloc - allocate memory
3753 * @size: how many bytes of memory are required.
3754 * @flags: the type of memory to allocate (see kmalloc).
3755 * @caller: function caller for debug tracking of the caller
3757 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3760 struct kmem_cache
*cachep
;
3763 /* If you want to save a few bytes .text space: replace
3765 * Then kmalloc uses the uninlined functions instead of the inline
3768 cachep
= __find_general_cachep(size
, flags
);
3769 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3771 ret
= __cache_alloc(cachep
, flags
, caller
);
3773 trace_kmalloc((unsigned long) caller
, ret
,
3774 size
, cachep
->size
, flags
);
3780 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3781 void *__kmalloc(size_t size
, gfp_t flags
)
3783 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3785 EXPORT_SYMBOL(__kmalloc
);
3787 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3789 return __do_kmalloc(size
, flags
, (void *)caller
);
3791 EXPORT_SYMBOL(__kmalloc_track_caller
);
3794 void *__kmalloc(size_t size
, gfp_t flags
)
3796 return __do_kmalloc(size
, flags
, NULL
);
3798 EXPORT_SYMBOL(__kmalloc
);
3802 * kmem_cache_free - Deallocate an object
3803 * @cachep: The cache the allocation was from.
3804 * @objp: The previously allocated object.
3806 * Free an object which was previously allocated from this
3809 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3811 unsigned long flags
;
3813 local_irq_save(flags
);
3814 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3815 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3816 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3817 __cache_free(cachep
, objp
, __builtin_return_address(0));
3818 local_irq_restore(flags
);
3820 trace_kmem_cache_free(_RET_IP_
, objp
);
3822 EXPORT_SYMBOL(kmem_cache_free
);
3825 * kfree - free previously allocated memory
3826 * @objp: pointer returned by kmalloc.
3828 * If @objp is NULL, no operation is performed.
3830 * Don't free memory not originally allocated by kmalloc()
3831 * or you will run into trouble.
3833 void kfree(const void *objp
)
3835 struct kmem_cache
*c
;
3836 unsigned long flags
;
3838 trace_kfree(_RET_IP_
, objp
);
3840 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3842 local_irq_save(flags
);
3843 kfree_debugcheck(objp
);
3844 c
= virt_to_cache(objp
);
3845 debug_check_no_locks_freed(objp
, c
->object_size
);
3847 debug_check_no_obj_freed(objp
, c
->object_size
);
3848 __cache_free(c
, (void *)objp
, __builtin_return_address(0));
3849 local_irq_restore(flags
);
3851 EXPORT_SYMBOL(kfree
);
3853 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3855 return cachep
->object_size
;
3857 EXPORT_SYMBOL(kmem_cache_size
);
3860 * This initializes kmem_list3 or resizes various caches for all nodes.
3862 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3865 struct kmem_list3
*l3
;
3866 struct array_cache
*new_shared
;
3867 struct array_cache
**new_alien
= NULL
;
3869 for_each_online_node(node
) {
3871 if (use_alien_caches
) {
3872 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3878 if (cachep
->shared
) {
3879 new_shared
= alloc_arraycache(node
,
3880 cachep
->shared
*cachep
->batchcount
,
3883 free_alien_cache(new_alien
);
3888 l3
= cachep
->nodelists
[node
];
3890 struct array_cache
*shared
= l3
->shared
;
3892 spin_lock_irq(&l3
->list_lock
);
3895 free_block(cachep
, shared
->entry
,
3896 shared
->avail
, node
);
3898 l3
->shared
= new_shared
;
3900 l3
->alien
= new_alien
;
3903 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3904 cachep
->batchcount
+ cachep
->num
;
3905 spin_unlock_irq(&l3
->list_lock
);
3907 free_alien_cache(new_alien
);
3910 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3912 free_alien_cache(new_alien
);
3917 kmem_list3_init(l3
);
3918 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3919 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3920 l3
->shared
= new_shared
;
3921 l3
->alien
= new_alien
;
3922 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3923 cachep
->batchcount
+ cachep
->num
;
3924 cachep
->nodelists
[node
] = l3
;
3929 if (!cachep
->list
.next
) {
3930 /* Cache is not active yet. Roll back what we did */
3933 if (cachep
->nodelists
[node
]) {
3934 l3
= cachep
->nodelists
[node
];
3937 free_alien_cache(l3
->alien
);
3939 cachep
->nodelists
[node
] = NULL
;
3947 struct ccupdate_struct
{
3948 struct kmem_cache
*cachep
;
3949 struct array_cache
*new[0];
3952 static void do_ccupdate_local(void *info
)
3954 struct ccupdate_struct
*new = info
;
3955 struct array_cache
*old
;
3958 old
= cpu_cache_get(new->cachep
);
3960 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3961 new->new[smp_processor_id()] = old
;
3964 /* Always called with the slab_mutex held */
3965 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3966 int batchcount
, int shared
, gfp_t gfp
)
3968 struct ccupdate_struct
*new;
3971 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3976 for_each_online_cpu(i
) {
3977 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3980 for (i
--; i
>= 0; i
--)
3986 new->cachep
= cachep
;
3988 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3991 cachep
->batchcount
= batchcount
;
3992 cachep
->limit
= limit
;
3993 cachep
->shared
= shared
;
3995 for_each_online_cpu(i
) {
3996 struct array_cache
*ccold
= new->new[i
];
3999 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4000 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4001 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4005 return alloc_kmemlist(cachep
, gfp
);
4008 /* Called with slab_mutex held always */
4009 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4015 * The head array serves three purposes:
4016 * - create a LIFO ordering, i.e. return objects that are cache-warm
4017 * - reduce the number of spinlock operations.
4018 * - reduce the number of linked list operations on the slab and
4019 * bufctl chains: array operations are cheaper.
4020 * The numbers are guessed, we should auto-tune as described by
4023 if (cachep
->size
> 131072)
4025 else if (cachep
->size
> PAGE_SIZE
)
4027 else if (cachep
->size
> 1024)
4029 else if (cachep
->size
> 256)
4035 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4036 * allocation behaviour: Most allocs on one cpu, most free operations
4037 * on another cpu. For these cases, an efficient object passing between
4038 * cpus is necessary. This is provided by a shared array. The array
4039 * replaces Bonwick's magazine layer.
4040 * On uniprocessor, it's functionally equivalent (but less efficient)
4041 * to a larger limit. Thus disabled by default.
4044 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4049 * With debugging enabled, large batchcount lead to excessively long
4050 * periods with disabled local interrupts. Limit the batchcount
4055 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4057 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4058 cachep
->name
, -err
);
4063 * Drain an array if it contains any elements taking the l3 lock only if
4064 * necessary. Note that the l3 listlock also protects the array_cache
4065 * if drain_array() is used on the shared array.
4067 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4068 struct array_cache
*ac
, int force
, int node
)
4072 if (!ac
|| !ac
->avail
)
4074 if (ac
->touched
&& !force
) {
4077 spin_lock_irq(&l3
->list_lock
);
4079 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4080 if (tofree
> ac
->avail
)
4081 tofree
= (ac
->avail
+ 1) / 2;
4082 free_block(cachep
, ac
->entry
, tofree
, node
);
4083 ac
->avail
-= tofree
;
4084 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4085 sizeof(void *) * ac
->avail
);
4087 spin_unlock_irq(&l3
->list_lock
);
4092 * cache_reap - Reclaim memory from caches.
4093 * @w: work descriptor
4095 * Called from workqueue/eventd every few seconds.
4097 * - clear the per-cpu caches for this CPU.
4098 * - return freeable pages to the main free memory pool.
4100 * If we cannot acquire the cache chain mutex then just give up - we'll try
4101 * again on the next iteration.
4103 static void cache_reap(struct work_struct
*w
)
4105 struct kmem_cache
*searchp
;
4106 struct kmem_list3
*l3
;
4107 int node
= numa_mem_id();
4108 struct delayed_work
*work
= to_delayed_work(w
);
4110 if (!mutex_trylock(&slab_mutex
))
4111 /* Give up. Setup the next iteration. */
4114 list_for_each_entry(searchp
, &slab_caches
, list
) {
4118 * We only take the l3 lock if absolutely necessary and we
4119 * have established with reasonable certainty that
4120 * we can do some work if the lock was obtained.
4122 l3
= searchp
->nodelists
[node
];
4124 reap_alien(searchp
, l3
);
4126 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4129 * These are racy checks but it does not matter
4130 * if we skip one check or scan twice.
4132 if (time_after(l3
->next_reap
, jiffies
))
4135 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4137 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4139 if (l3
->free_touched
)
4140 l3
->free_touched
= 0;
4144 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4145 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4146 STATS_ADD_REAPED(searchp
, freed
);
4152 mutex_unlock(&slab_mutex
);
4155 /* Set up the next iteration */
4156 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4159 #ifdef CONFIG_SLABINFO
4161 static void print_slabinfo_header(struct seq_file
*m
)
4164 * Output format version, so at least we can change it
4165 * without _too_ many complaints.
4168 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4170 seq_puts(m
, "slabinfo - version: 2.1\n");
4172 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4173 "<objperslab> <pagesperslab>");
4174 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4175 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4177 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4178 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4179 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4184 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4188 mutex_lock(&slab_mutex
);
4190 print_slabinfo_header(m
);
4192 return seq_list_start(&slab_caches
, *pos
);
4195 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4197 return seq_list_next(p
, &slab_caches
, pos
);
4200 static void s_stop(struct seq_file
*m
, void *p
)
4202 mutex_unlock(&slab_mutex
);
4205 static int s_show(struct seq_file
*m
, void *p
)
4207 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4209 unsigned long active_objs
;
4210 unsigned long num_objs
;
4211 unsigned long active_slabs
= 0;
4212 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4216 struct kmem_list3
*l3
;
4220 for_each_online_node(node
) {
4221 l3
= cachep
->nodelists
[node
];
4226 spin_lock_irq(&l3
->list_lock
);
4228 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4229 if (slabp
->inuse
!= cachep
->num
&& !error
)
4230 error
= "slabs_full accounting error";
4231 active_objs
+= cachep
->num
;
4234 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4235 if (slabp
->inuse
== cachep
->num
&& !error
)
4236 error
= "slabs_partial inuse accounting error";
4237 if (!slabp
->inuse
&& !error
)
4238 error
= "slabs_partial/inuse accounting error";
4239 active_objs
+= slabp
->inuse
;
4242 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4243 if (slabp
->inuse
&& !error
)
4244 error
= "slabs_free/inuse accounting error";
4247 free_objects
+= l3
->free_objects
;
4249 shared_avail
+= l3
->shared
->avail
;
4251 spin_unlock_irq(&l3
->list_lock
);
4253 num_slabs
+= active_slabs
;
4254 num_objs
= num_slabs
* cachep
->num
;
4255 if (num_objs
- active_objs
!= free_objects
&& !error
)
4256 error
= "free_objects accounting error";
4258 name
= cachep
->name
;
4260 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4262 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4263 name
, active_objs
, num_objs
, cachep
->size
,
4264 cachep
->num
, (1 << cachep
->gfporder
));
4265 seq_printf(m
, " : tunables %4u %4u %4u",
4266 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4267 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4268 active_slabs
, num_slabs
, shared_avail
);
4271 unsigned long high
= cachep
->high_mark
;
4272 unsigned long allocs
= cachep
->num_allocations
;
4273 unsigned long grown
= cachep
->grown
;
4274 unsigned long reaped
= cachep
->reaped
;
4275 unsigned long errors
= cachep
->errors
;
4276 unsigned long max_freeable
= cachep
->max_freeable
;
4277 unsigned long node_allocs
= cachep
->node_allocs
;
4278 unsigned long node_frees
= cachep
->node_frees
;
4279 unsigned long overflows
= cachep
->node_overflow
;
4281 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4282 "%4lu %4lu %4lu %4lu %4lu",
4283 allocs
, high
, grown
,
4284 reaped
, errors
, max_freeable
, node_allocs
,
4285 node_frees
, overflows
);
4289 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4290 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4291 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4292 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4294 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4295 allochit
, allocmiss
, freehit
, freemiss
);
4303 * slabinfo_op - iterator that generates /proc/slabinfo
4312 * num-pages-per-slab
4313 * + further values on SMP and with statistics enabled
4316 static const struct seq_operations slabinfo_op
= {
4323 #define MAX_SLABINFO_WRITE 128
4325 * slabinfo_write - Tuning for the slab allocator
4327 * @buffer: user buffer
4328 * @count: data length
4331 static ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4332 size_t count
, loff_t
*ppos
)
4334 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4335 int limit
, batchcount
, shared
, res
;
4336 struct kmem_cache
*cachep
;
4338 if (count
> MAX_SLABINFO_WRITE
)
4340 if (copy_from_user(&kbuf
, buffer
, count
))
4342 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4344 tmp
= strchr(kbuf
, ' ');
4349 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4352 /* Find the cache in the chain of caches. */
4353 mutex_lock(&slab_mutex
);
4355 list_for_each_entry(cachep
, &slab_caches
, list
) {
4356 if (!strcmp(cachep
->name
, kbuf
)) {
4357 if (limit
< 1 || batchcount
< 1 ||
4358 batchcount
> limit
|| shared
< 0) {
4361 res
= do_tune_cpucache(cachep
, limit
,
4368 mutex_unlock(&slab_mutex
);
4374 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4376 return seq_open(file
, &slabinfo_op
);
4379 static const struct file_operations proc_slabinfo_operations
= {
4380 .open
= slabinfo_open
,
4382 .write
= slabinfo_write
,
4383 .llseek
= seq_lseek
,
4384 .release
= seq_release
,
4387 #ifdef CONFIG_DEBUG_SLAB_LEAK
4389 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4391 mutex_lock(&slab_mutex
);
4392 return seq_list_start(&slab_caches
, *pos
);
4395 static inline int add_caller(unsigned long *n
, unsigned long v
)
4405 unsigned long *q
= p
+ 2 * i
;
4419 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4425 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4431 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4432 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4434 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4439 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4441 #ifdef CONFIG_KALLSYMS
4442 unsigned long offset
, size
;
4443 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4445 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4446 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4448 seq_printf(m
, " [%s]", modname
);
4452 seq_printf(m
, "%p", (void *)address
);
4455 static int leaks_show(struct seq_file
*m
, void *p
)
4457 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4459 struct kmem_list3
*l3
;
4461 unsigned long *n
= m
->private;
4465 if (!(cachep
->flags
& SLAB_STORE_USER
))
4467 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4470 /* OK, we can do it */
4474 for_each_online_node(node
) {
4475 l3
= cachep
->nodelists
[node
];
4480 spin_lock_irq(&l3
->list_lock
);
4482 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4483 handle_slab(n
, cachep
, slabp
);
4484 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4485 handle_slab(n
, cachep
, slabp
);
4486 spin_unlock_irq(&l3
->list_lock
);
4488 name
= cachep
->name
;
4490 /* Increase the buffer size */
4491 mutex_unlock(&slab_mutex
);
4492 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4494 /* Too bad, we are really out */
4496 mutex_lock(&slab_mutex
);
4499 *(unsigned long *)m
->private = n
[0] * 2;
4501 mutex_lock(&slab_mutex
);
4502 /* Now make sure this entry will be retried */
4506 for (i
= 0; i
< n
[1]; i
++) {
4507 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4508 show_symbol(m
, n
[2*i
+2]);
4515 static const struct seq_operations slabstats_op
= {
4516 .start
= leaks_start
,
4522 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4524 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4527 ret
= seq_open(file
, &slabstats_op
);
4529 struct seq_file
*m
= file
->private_data
;
4530 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4539 static const struct file_operations proc_slabstats_operations
= {
4540 .open
= slabstats_open
,
4542 .llseek
= seq_lseek
,
4543 .release
= seq_release_private
,
4547 static int __init
slab_proc_init(void)
4549 proc_create("slabinfo",S_IWUSR
|S_IRUSR
,NULL
,&proc_slabinfo_operations
);
4550 #ifdef CONFIG_DEBUG_SLAB_LEAK
4551 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4555 module_init(slab_proc_init
);
4559 * ksize - get the actual amount of memory allocated for a given object
4560 * @objp: Pointer to the object
4562 * kmalloc may internally round up allocations and return more memory
4563 * than requested. ksize() can be used to determine the actual amount of
4564 * memory allocated. The caller may use this additional memory, even though
4565 * a smaller amount of memory was initially specified with the kmalloc call.
4566 * The caller must guarantee that objp points to a valid object previously
4567 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4568 * must not be freed during the duration of the call.
4570 size_t ksize(const void *objp
)
4573 if (unlikely(objp
== ZERO_SIZE_PTR
))
4576 return virt_to_cache(objp
)->object_size
;
4578 EXPORT_SYMBOL(ksize
);