3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
121 #include <net/sock.h>
123 #include <asm/cacheflush.h>
124 #include <asm/tlbflush.h>
125 #include <asm/page.h>
127 #include <trace/events/kmem.h>
129 #include "internal.h"
132 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
133 * 0 for faster, smaller code (especially in the critical paths).
135 * STATS - 1 to collect stats for /proc/slabinfo.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 #ifdef CONFIG_DEBUG_SLAB
144 #define FORCED_DEBUG 1
148 #define FORCED_DEBUG 0
151 /* Shouldn't this be in a header file somewhere? */
152 #define BYTES_PER_WORD sizeof(void *)
153 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155 #ifndef ARCH_KMALLOC_FLAGS
156 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 * true if a page was allocated from pfmemalloc reserves for network-based
163 static bool pfmemalloc_active __read_mostly
;
168 * Bufctl's are used for linking objs within a slab
171 * This implementation relies on "struct page" for locating the cache &
172 * slab an object belongs to.
173 * This allows the bufctl structure to be small (one int), but limits
174 * the number of objects a slab (not a cache) can contain when off-slab
175 * bufctls are used. The limit is the size of the largest general cache
176 * that does not use off-slab slabs.
177 * For 32bit archs with 4 kB pages, is this 56.
178 * This is not serious, as it is only for large objects, when it is unwise
179 * to have too many per slab.
180 * Note: This limit can be raised by introducing a general cache whose size
181 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
184 typedef unsigned int kmem_bufctl_t
;
185 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
186 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
187 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
188 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
193 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
194 * arrange for kmem_freepages to be called via RCU. This is useful if
195 * we need to approach a kernel structure obliquely, from its address
196 * obtained without the usual locking. We can lock the structure to
197 * stabilize it and check it's still at the given address, only if we
198 * can be sure that the memory has not been meanwhile reused for some
199 * other kind of object (which our subsystem's lock might corrupt).
201 * rcu_read_lock before reading the address, then rcu_read_unlock after
202 * taking the spinlock within the structure expected at that address.
205 struct rcu_head head
;
206 struct kmem_cache
*cachep
;
213 * Manages the objs in a slab. Placed either at the beginning of mem allocated
214 * for a slab, or allocated from an general cache.
215 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list
;
221 unsigned long colouroff
;
222 void *s_mem
; /* including colour offset */
223 unsigned int inuse
; /* num of objs active in slab */
225 unsigned short nodeid
;
227 struct slab_rcu __slab_cover_slab_rcu
;
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
239 * The limit is stored in the per-cpu structure to reduce the data cache
246 unsigned int batchcount
;
247 unsigned int touched
;
250 * Must have this definition in here for the proper
251 * alignment of array_cache. Also simplifies accessing
254 * Entries should not be directly dereferenced as
255 * entries belonging to slabs marked pfmemalloc will
256 * have the lower bits set SLAB_OBJ_PFMEMALLOC
260 #define SLAB_OBJ_PFMEMALLOC 1
261 static inline bool is_obj_pfmemalloc(void *objp
)
263 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
266 static inline void set_obj_pfmemalloc(void **objp
)
268 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
272 static inline void clear_obj_pfmemalloc(void **objp
)
274 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init
{
283 struct array_cache cache
;
284 void *entries
[BOOT_CPUCACHE_ENTRIES
];
288 * The slab lists for all objects.
291 struct list_head slabs_partial
; /* partial list first, better asm code */
292 struct list_head slabs_full
;
293 struct list_head slabs_free
;
294 unsigned long free_objects
;
295 unsigned int free_limit
;
296 unsigned int colour_next
; /* Per-node cache coloring */
297 spinlock_t list_lock
;
298 struct array_cache
*shared
; /* shared per node */
299 struct array_cache
**alien
; /* on other nodes */
300 unsigned long next_reap
; /* updated without locking */
301 int free_touched
; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
308 static struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
309 #define CACHE_CACHE 0
310 #define SIZE_AC MAX_NUMNODES
311 #define SIZE_L3 (2 * MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache
*cache
,
314 struct kmem_list3
*l3
, int tofree
);
315 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
317 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
318 static void cache_reap(struct work_struct
*unused
);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline
int index_of(const size_t size
)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size
)) {
336 #include <linux/kmalloc_sizes.h>
344 static int slab_early_init
= 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3
*parent
)
351 INIT_LIST_HEAD(&parent
->slabs_full
);
352 INIT_LIST_HEAD(&parent
->slabs_partial
);
353 INIT_LIST_HEAD(&parent
->slabs_free
);
354 parent
->shared
= NULL
;
355 parent
->alien
= NULL
;
356 parent
->colour_next
= 0;
357 spin_lock_init(&parent
->list_lock
);
358 parent
->free_objects
= 0;
359 parent
->free_touched
= 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
375 #define CFLGS_OFF_SLAB (0x80000000UL)
376 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
378 #define BATCHREFILL_LIMIT 16
380 * Optimization question: fewer reaps means less probability for unnessary
381 * cpucache drain/refill cycles.
383 * OTOH the cpuarrays can contain lots of objects,
384 * which could lock up otherwise freeable slabs.
386 #define REAPTIMEOUT_CPUC (2*HZ)
387 #define REAPTIMEOUT_LIST3 (4*HZ)
390 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
391 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
392 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
393 #define STATS_INC_GROWN(x) ((x)->grown++)
394 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
395 #define STATS_SET_HIGH(x) \
397 if ((x)->num_active > (x)->high_mark) \
398 (x)->high_mark = (x)->num_active; \
400 #define STATS_INC_ERR(x) ((x)->errors++)
401 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
402 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
403 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
404 #define STATS_SET_FREEABLE(x, i) \
406 if ((x)->max_freeable < i) \
407 (x)->max_freeable = i; \
409 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
410 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
411 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
412 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
414 #define STATS_INC_ACTIVE(x) do { } while (0)
415 #define STATS_DEC_ACTIVE(x) do { } while (0)
416 #define STATS_INC_ALLOCED(x) do { } while (0)
417 #define STATS_INC_GROWN(x) do { } while (0)
418 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
419 #define STATS_SET_HIGH(x) do { } while (0)
420 #define STATS_INC_ERR(x) do { } while (0)
421 #define STATS_INC_NODEALLOCS(x) do { } while (0)
422 #define STATS_INC_NODEFREES(x) do { } while (0)
423 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
424 #define STATS_SET_FREEABLE(x, i) do { } while (0)
425 #define STATS_INC_ALLOCHIT(x) do { } while (0)
426 #define STATS_INC_ALLOCMISS(x) do { } while (0)
427 #define STATS_INC_FREEHIT(x) do { } while (0)
428 #define STATS_INC_FREEMISS(x) do { } while (0)
434 * memory layout of objects:
436 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
437 * the end of an object is aligned with the end of the real
438 * allocation. Catches writes behind the end of the allocation.
439 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
441 * cachep->obj_offset: The real object.
442 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
443 * cachep->size - 1* BYTES_PER_WORD: last caller address
444 * [BYTES_PER_WORD long]
446 static int obj_offset(struct kmem_cache
*cachep
)
448 return cachep
->obj_offset
;
451 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
453 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
454 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
455 sizeof(unsigned long long));
458 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
460 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
461 if (cachep
->flags
& SLAB_STORE_USER
)
462 return (unsigned long long *)(objp
+ cachep
->size
-
463 sizeof(unsigned long long) -
465 return (unsigned long long *) (objp
+ cachep
->size
-
466 sizeof(unsigned long long));
469 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
471 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
472 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
477 #define obj_offset(x) 0
478 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
479 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
480 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
485 * Do not go above this order unless 0 objects fit into the slab or
486 * overridden on the command line.
488 #define SLAB_MAX_ORDER_HI 1
489 #define SLAB_MAX_ORDER_LO 0
490 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
491 static bool slab_max_order_set __initdata
;
493 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
495 struct page
*page
= virt_to_head_page(obj
);
496 return page
->slab_cache
;
499 static inline struct slab
*virt_to_slab(const void *obj
)
501 struct page
*page
= virt_to_head_page(obj
);
503 VM_BUG_ON(!PageSlab(page
));
504 return page
->slab_page
;
507 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
510 return slab
->s_mem
+ cache
->size
* idx
;
514 * We want to avoid an expensive divide : (offset / cache->size)
515 * Using the fact that size is a constant for a particular cache,
516 * we can replace (offset / cache->size) by
517 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
519 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
520 const struct slab
*slab
, void *obj
)
522 u32 offset
= (obj
- slab
->s_mem
);
523 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
527 * These are the default caches for kmalloc. Custom caches can have other sizes.
529 struct cache_sizes malloc_sizes
[] = {
530 #define CACHE(x) { .cs_size = (x) },
531 #include <linux/kmalloc_sizes.h>
535 EXPORT_SYMBOL(malloc_sizes
);
537 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
543 static struct cache_names __initdata cache_names
[] = {
544 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
545 #include <linux/kmalloc_sizes.h>
550 static struct arraycache_init initarray_generic
=
551 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
553 /* internal cache of cache description objs */
554 static struct kmem_cache kmem_cache_boot
= {
556 .limit
= BOOT_CPUCACHE_ENTRIES
,
558 .size
= sizeof(struct kmem_cache
),
559 .name
= "kmem_cache",
562 #define BAD_ALIEN_MAGIC 0x01020304ul
564 #ifdef CONFIG_LOCKDEP
567 * Slab sometimes uses the kmalloc slabs to store the slab headers
568 * for other slabs "off slab".
569 * The locking for this is tricky in that it nests within the locks
570 * of all other slabs in a few places; to deal with this special
571 * locking we put on-slab caches into a separate lock-class.
573 * We set lock class for alien array caches which are up during init.
574 * The lock annotation will be lost if all cpus of a node goes down and
575 * then comes back up during hotplug
577 static struct lock_class_key on_slab_l3_key
;
578 static struct lock_class_key on_slab_alc_key
;
580 static struct lock_class_key debugobj_l3_key
;
581 static struct lock_class_key debugobj_alc_key
;
583 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
584 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
587 struct array_cache
**alc
;
588 struct kmem_list3
*l3
;
591 l3
= cachep
->nodelists
[q
];
595 lockdep_set_class(&l3
->list_lock
, l3_key
);
598 * FIXME: This check for BAD_ALIEN_MAGIC
599 * should go away when common slab code is taught to
600 * work even without alien caches.
601 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
602 * for alloc_alien_cache,
604 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
608 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
612 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
614 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
617 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
621 for_each_online_node(node
)
622 slab_set_debugobj_lock_classes_node(cachep
, node
);
625 static void init_node_lock_keys(int q
)
627 struct cache_sizes
*s
= malloc_sizes
;
632 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
633 struct kmem_list3
*l3
;
635 l3
= s
->cs_cachep
->nodelists
[q
];
636 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
639 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
640 &on_slab_alc_key
, q
);
644 static inline void init_lock_keys(void)
649 init_node_lock_keys(node
);
652 static void init_node_lock_keys(int q
)
656 static inline void init_lock_keys(void)
660 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
664 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
669 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
671 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
673 return cachep
->array
[smp_processor_id()];
676 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
679 struct cache_sizes
*csizep
= malloc_sizes
;
682 /* This happens if someone tries to call
683 * kmem_cache_create(), or __kmalloc(), before
684 * the generic caches are initialized.
686 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
689 return ZERO_SIZE_PTR
;
691 while (size
> csizep
->cs_size
)
695 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
696 * has cs_{dma,}cachep==NULL. Thus no special case
697 * for large kmalloc calls required.
699 #ifdef CONFIG_ZONE_DMA
700 if (unlikely(gfpflags
& GFP_DMA
))
701 return csizep
->cs_dmacachep
;
703 return csizep
->cs_cachep
;
706 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
708 return __find_general_cachep(size
, gfpflags
);
711 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
713 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
717 * Calculate the number of objects and left-over bytes for a given buffer size.
719 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
720 size_t align
, int flags
, size_t *left_over
,
725 size_t slab_size
= PAGE_SIZE
<< gfporder
;
728 * The slab management structure can be either off the slab or
729 * on it. For the latter case, the memory allocated for a
733 * - One kmem_bufctl_t for each object
734 * - Padding to respect alignment of @align
735 * - @buffer_size bytes for each object
737 * If the slab management structure is off the slab, then the
738 * alignment will already be calculated into the size. Because
739 * the slabs are all pages aligned, the objects will be at the
740 * correct alignment when allocated.
742 if (flags
& CFLGS_OFF_SLAB
) {
744 nr_objs
= slab_size
/ buffer_size
;
746 if (nr_objs
> SLAB_LIMIT
)
747 nr_objs
= SLAB_LIMIT
;
750 * Ignore padding for the initial guess. The padding
751 * is at most @align-1 bytes, and @buffer_size is at
752 * least @align. In the worst case, this result will
753 * be one greater than the number of objects that fit
754 * into the memory allocation when taking the padding
757 nr_objs
= (slab_size
- sizeof(struct slab
)) /
758 (buffer_size
+ sizeof(kmem_bufctl_t
));
761 * This calculated number will be either the right
762 * amount, or one greater than what we want.
764 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
768 if (nr_objs
> SLAB_LIMIT
)
769 nr_objs
= SLAB_LIMIT
;
771 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
774 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
778 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
780 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
783 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
784 function
, cachep
->name
, msg
);
786 add_taint(TAINT_BAD_PAGE
);
791 * By default on NUMA we use alien caches to stage the freeing of
792 * objects allocated from other nodes. This causes massive memory
793 * inefficiencies when using fake NUMA setup to split memory into a
794 * large number of small nodes, so it can be disabled on the command
798 static int use_alien_caches __read_mostly
= 1;
799 static int __init
noaliencache_setup(char *s
)
801 use_alien_caches
= 0;
804 __setup("noaliencache", noaliencache_setup
);
806 static int __init
slab_max_order_setup(char *str
)
808 get_option(&str
, &slab_max_order
);
809 slab_max_order
= slab_max_order
< 0 ? 0 :
810 min(slab_max_order
, MAX_ORDER
- 1);
811 slab_max_order_set
= true;
815 __setup("slab_max_order=", slab_max_order_setup
);
819 * Special reaping functions for NUMA systems called from cache_reap().
820 * These take care of doing round robin flushing of alien caches (containing
821 * objects freed on different nodes from which they were allocated) and the
822 * flushing of remote pcps by calling drain_node_pages.
824 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
826 static void init_reap_node(int cpu
)
830 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
831 if (node
== MAX_NUMNODES
)
832 node
= first_node(node_online_map
);
834 per_cpu(slab_reap_node
, cpu
) = node
;
837 static void next_reap_node(void)
839 int node
= __this_cpu_read(slab_reap_node
);
841 node
= next_node(node
, node_online_map
);
842 if (unlikely(node
>= MAX_NUMNODES
))
843 node
= first_node(node_online_map
);
844 __this_cpu_write(slab_reap_node
, node
);
848 #define init_reap_node(cpu) do { } while (0)
849 #define next_reap_node(void) do { } while (0)
853 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
854 * via the workqueue/eventd.
855 * Add the CPU number into the expiration time to minimize the possibility of
856 * the CPUs getting into lockstep and contending for the global cache chain
859 static void __cpuinit
start_cpu_timer(int cpu
)
861 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
864 * When this gets called from do_initcalls via cpucache_init(),
865 * init_workqueues() has already run, so keventd will be setup
868 if (keventd_up() && reap_work
->work
.func
== NULL
) {
870 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
871 schedule_delayed_work_on(cpu
, reap_work
,
872 __round_jiffies_relative(HZ
, cpu
));
876 static struct array_cache
*alloc_arraycache(int node
, int entries
,
877 int batchcount
, gfp_t gfp
)
879 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
880 struct array_cache
*nc
= NULL
;
882 nc
= kmalloc_node(memsize
, gfp
, node
);
884 * The array_cache structures contain pointers to free object.
885 * However, when such objects are allocated or transferred to another
886 * cache the pointers are not cleared and they could be counted as
887 * valid references during a kmemleak scan. Therefore, kmemleak must
888 * not scan such objects.
890 kmemleak_no_scan(nc
);
894 nc
->batchcount
= batchcount
;
896 spin_lock_init(&nc
->lock
);
901 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
903 struct page
*page
= virt_to_page(slabp
->s_mem
);
905 return PageSlabPfmemalloc(page
);
908 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
909 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
910 struct array_cache
*ac
)
912 struct kmem_list3
*l3
= cachep
->nodelists
[numa_mem_id()];
916 if (!pfmemalloc_active
)
919 spin_lock_irqsave(&l3
->list_lock
, flags
);
920 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
921 if (is_slab_pfmemalloc(slabp
))
924 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
925 if (is_slab_pfmemalloc(slabp
))
928 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
929 if (is_slab_pfmemalloc(slabp
))
932 pfmemalloc_active
= false;
934 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
937 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
938 gfp_t flags
, bool force_refill
)
941 void *objp
= ac
->entry
[--ac
->avail
];
943 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
944 if (unlikely(is_obj_pfmemalloc(objp
))) {
945 struct kmem_list3
*l3
;
947 if (gfp_pfmemalloc_allowed(flags
)) {
948 clear_obj_pfmemalloc(&objp
);
952 /* The caller cannot use PFMEMALLOC objects, find another one */
953 for (i
= 0; i
< ac
->avail
; i
++) {
954 /* If a !PFMEMALLOC object is found, swap them */
955 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
957 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
958 ac
->entry
[ac
->avail
] = objp
;
964 * If there are empty slabs on the slabs_free list and we are
965 * being forced to refill the cache, mark this one !pfmemalloc.
967 l3
= cachep
->nodelists
[numa_mem_id()];
968 if (!list_empty(&l3
->slabs_free
) && force_refill
) {
969 struct slab
*slabp
= virt_to_slab(objp
);
970 ClearPageSlabPfmemalloc(virt_to_head_page(slabp
->s_mem
));
971 clear_obj_pfmemalloc(&objp
);
972 recheck_pfmemalloc_active(cachep
, ac
);
976 /* No !PFMEMALLOC objects available */
984 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
985 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
989 if (unlikely(sk_memalloc_socks()))
990 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
992 objp
= ac
->entry
[--ac
->avail
];
997 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1000 if (unlikely(pfmemalloc_active
)) {
1001 /* Some pfmemalloc slabs exist, check if this is one */
1002 struct page
*page
= virt_to_head_page(objp
);
1003 if (PageSlabPfmemalloc(page
))
1004 set_obj_pfmemalloc(&objp
);
1010 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
1013 if (unlikely(sk_memalloc_socks()))
1014 objp
= __ac_put_obj(cachep
, ac
, objp
);
1016 ac
->entry
[ac
->avail
++] = objp
;
1020 * Transfer objects in one arraycache to another.
1021 * Locking must be handled by the caller.
1023 * Return the number of entries transferred.
1025 static int transfer_objects(struct array_cache
*to
,
1026 struct array_cache
*from
, unsigned int max
)
1028 /* Figure out how many entries to transfer */
1029 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
1034 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1035 sizeof(void *) *nr
);
1044 #define drain_alien_cache(cachep, alien) do { } while (0)
1045 #define reap_alien(cachep, l3) do { } while (0)
1047 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1049 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1052 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1056 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1061 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1067 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1068 gfp_t flags
, int nodeid
)
1073 #else /* CONFIG_NUMA */
1075 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1076 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1078 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1080 struct array_cache
**ac_ptr
;
1081 int memsize
= sizeof(void *) * nr_node_ids
;
1086 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1089 if (i
== node
|| !node_online(i
))
1091 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1093 for (i
--; i
>= 0; i
--)
1103 static void free_alien_cache(struct array_cache
**ac_ptr
)
1114 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1115 struct array_cache
*ac
, int node
)
1117 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1120 spin_lock(&rl3
->list_lock
);
1122 * Stuff objects into the remote nodes shared array first.
1123 * That way we could avoid the overhead of putting the objects
1124 * into the free lists and getting them back later.
1127 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1129 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1131 spin_unlock(&rl3
->list_lock
);
1136 * Called from cache_reap() to regularly drain alien caches round robin.
1138 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1140 int node
= __this_cpu_read(slab_reap_node
);
1143 struct array_cache
*ac
= l3
->alien
[node
];
1145 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1146 __drain_alien_cache(cachep
, ac
, node
);
1147 spin_unlock_irq(&ac
->lock
);
1152 static void drain_alien_cache(struct kmem_cache
*cachep
,
1153 struct array_cache
**alien
)
1156 struct array_cache
*ac
;
1157 unsigned long flags
;
1159 for_each_online_node(i
) {
1162 spin_lock_irqsave(&ac
->lock
, flags
);
1163 __drain_alien_cache(cachep
, ac
, i
);
1164 spin_unlock_irqrestore(&ac
->lock
, flags
);
1169 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1171 struct slab
*slabp
= virt_to_slab(objp
);
1172 int nodeid
= slabp
->nodeid
;
1173 struct kmem_list3
*l3
;
1174 struct array_cache
*alien
= NULL
;
1177 node
= numa_mem_id();
1180 * Make sure we are not freeing a object from another node to the array
1181 * cache on this cpu.
1183 if (likely(slabp
->nodeid
== node
))
1186 l3
= cachep
->nodelists
[node
];
1187 STATS_INC_NODEFREES(cachep
);
1188 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1189 alien
= l3
->alien
[nodeid
];
1190 spin_lock(&alien
->lock
);
1191 if (unlikely(alien
->avail
== alien
->limit
)) {
1192 STATS_INC_ACOVERFLOW(cachep
);
1193 __drain_alien_cache(cachep
, alien
, nodeid
);
1195 ac_put_obj(cachep
, alien
, objp
);
1196 spin_unlock(&alien
->lock
);
1198 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1199 free_block(cachep
, &objp
, 1, nodeid
);
1200 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1207 * Allocates and initializes nodelists for a node on each slab cache, used for
1208 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1209 * will be allocated off-node since memory is not yet online for the new node.
1210 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1213 * Must hold slab_mutex.
1215 static int init_cache_nodelists_node(int node
)
1217 struct kmem_cache
*cachep
;
1218 struct kmem_list3
*l3
;
1219 const int memsize
= sizeof(struct kmem_list3
);
1221 list_for_each_entry(cachep
, &slab_caches
, list
) {
1223 * Set up the size64 kmemlist for cpu before we can
1224 * begin anything. Make sure some other cpu on this
1225 * node has not already allocated this
1227 if (!cachep
->nodelists
[node
]) {
1228 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1231 kmem_list3_init(l3
);
1232 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1233 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1236 * The l3s don't come and go as CPUs come and
1237 * go. slab_mutex is sufficient
1240 cachep
->nodelists
[node
] = l3
;
1243 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1244 cachep
->nodelists
[node
]->free_limit
=
1245 (1 + nr_cpus_node(node
)) *
1246 cachep
->batchcount
+ cachep
->num
;
1247 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1252 static void __cpuinit
cpuup_canceled(long cpu
)
1254 struct kmem_cache
*cachep
;
1255 struct kmem_list3
*l3
= NULL
;
1256 int node
= cpu_to_mem(cpu
);
1257 const struct cpumask
*mask
= cpumask_of_node(node
);
1259 list_for_each_entry(cachep
, &slab_caches
, list
) {
1260 struct array_cache
*nc
;
1261 struct array_cache
*shared
;
1262 struct array_cache
**alien
;
1264 /* cpu is dead; no one can alloc from it. */
1265 nc
= cachep
->array
[cpu
];
1266 cachep
->array
[cpu
] = NULL
;
1267 l3
= cachep
->nodelists
[node
];
1270 goto free_array_cache
;
1272 spin_lock_irq(&l3
->list_lock
);
1274 /* Free limit for this kmem_list3 */
1275 l3
->free_limit
-= cachep
->batchcount
;
1277 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1279 if (!cpumask_empty(mask
)) {
1280 spin_unlock_irq(&l3
->list_lock
);
1281 goto free_array_cache
;
1284 shared
= l3
->shared
;
1286 free_block(cachep
, shared
->entry
,
1287 shared
->avail
, node
);
1294 spin_unlock_irq(&l3
->list_lock
);
1298 drain_alien_cache(cachep
, alien
);
1299 free_alien_cache(alien
);
1305 * In the previous loop, all the objects were freed to
1306 * the respective cache's slabs, now we can go ahead and
1307 * shrink each nodelist to its limit.
1309 list_for_each_entry(cachep
, &slab_caches
, list
) {
1310 l3
= cachep
->nodelists
[node
];
1313 drain_freelist(cachep
, l3
, l3
->free_objects
);
1317 static int __cpuinit
cpuup_prepare(long cpu
)
1319 struct kmem_cache
*cachep
;
1320 struct kmem_list3
*l3
= NULL
;
1321 int node
= cpu_to_mem(cpu
);
1325 * We need to do this right in the beginning since
1326 * alloc_arraycache's are going to use this list.
1327 * kmalloc_node allows us to add the slab to the right
1328 * kmem_list3 and not this cpu's kmem_list3
1330 err
= init_cache_nodelists_node(node
);
1335 * Now we can go ahead with allocating the shared arrays and
1338 list_for_each_entry(cachep
, &slab_caches
, list
) {
1339 struct array_cache
*nc
;
1340 struct array_cache
*shared
= NULL
;
1341 struct array_cache
**alien
= NULL
;
1343 nc
= alloc_arraycache(node
, cachep
->limit
,
1344 cachep
->batchcount
, GFP_KERNEL
);
1347 if (cachep
->shared
) {
1348 shared
= alloc_arraycache(node
,
1349 cachep
->shared
* cachep
->batchcount
,
1350 0xbaadf00d, GFP_KERNEL
);
1356 if (use_alien_caches
) {
1357 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1364 cachep
->array
[cpu
] = nc
;
1365 l3
= cachep
->nodelists
[node
];
1368 spin_lock_irq(&l3
->list_lock
);
1371 * We are serialised from CPU_DEAD or
1372 * CPU_UP_CANCELLED by the cpucontrol lock
1374 l3
->shared
= shared
;
1383 spin_unlock_irq(&l3
->list_lock
);
1385 free_alien_cache(alien
);
1386 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1387 slab_set_debugobj_lock_classes_node(cachep
, node
);
1389 init_node_lock_keys(node
);
1393 cpuup_canceled(cpu
);
1397 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1398 unsigned long action
, void *hcpu
)
1400 long cpu
= (long)hcpu
;
1404 case CPU_UP_PREPARE
:
1405 case CPU_UP_PREPARE_FROZEN
:
1406 mutex_lock(&slab_mutex
);
1407 err
= cpuup_prepare(cpu
);
1408 mutex_unlock(&slab_mutex
);
1411 case CPU_ONLINE_FROZEN
:
1412 start_cpu_timer(cpu
);
1414 #ifdef CONFIG_HOTPLUG_CPU
1415 case CPU_DOWN_PREPARE
:
1416 case CPU_DOWN_PREPARE_FROZEN
:
1418 * Shutdown cache reaper. Note that the slab_mutex is
1419 * held so that if cache_reap() is invoked it cannot do
1420 * anything expensive but will only modify reap_work
1421 * and reschedule the timer.
1423 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1424 /* Now the cache_reaper is guaranteed to be not running. */
1425 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1427 case CPU_DOWN_FAILED
:
1428 case CPU_DOWN_FAILED_FROZEN
:
1429 start_cpu_timer(cpu
);
1432 case CPU_DEAD_FROZEN
:
1434 * Even if all the cpus of a node are down, we don't free the
1435 * kmem_list3 of any cache. This to avoid a race between
1436 * cpu_down, and a kmalloc allocation from another cpu for
1437 * memory from the node of the cpu going down. The list3
1438 * structure is usually allocated from kmem_cache_create() and
1439 * gets destroyed at kmem_cache_destroy().
1443 case CPU_UP_CANCELED
:
1444 case CPU_UP_CANCELED_FROZEN
:
1445 mutex_lock(&slab_mutex
);
1446 cpuup_canceled(cpu
);
1447 mutex_unlock(&slab_mutex
);
1450 return notifier_from_errno(err
);
1453 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1454 &cpuup_callback
, NULL
, 0
1457 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1459 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1460 * Returns -EBUSY if all objects cannot be drained so that the node is not
1463 * Must hold slab_mutex.
1465 static int __meminit
drain_cache_nodelists_node(int node
)
1467 struct kmem_cache
*cachep
;
1470 list_for_each_entry(cachep
, &slab_caches
, list
) {
1471 struct kmem_list3
*l3
;
1473 l3
= cachep
->nodelists
[node
];
1477 drain_freelist(cachep
, l3
, l3
->free_objects
);
1479 if (!list_empty(&l3
->slabs_full
) ||
1480 !list_empty(&l3
->slabs_partial
)) {
1488 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1489 unsigned long action
, void *arg
)
1491 struct memory_notify
*mnb
= arg
;
1495 nid
= mnb
->status_change_nid
;
1500 case MEM_GOING_ONLINE
:
1501 mutex_lock(&slab_mutex
);
1502 ret
= init_cache_nodelists_node(nid
);
1503 mutex_unlock(&slab_mutex
);
1505 case MEM_GOING_OFFLINE
:
1506 mutex_lock(&slab_mutex
);
1507 ret
= drain_cache_nodelists_node(nid
);
1508 mutex_unlock(&slab_mutex
);
1512 case MEM_CANCEL_ONLINE
:
1513 case MEM_CANCEL_OFFLINE
:
1517 return notifier_from_errno(ret
);
1519 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1522 * swap the static kmem_list3 with kmalloced memory
1524 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1527 struct kmem_list3
*ptr
;
1529 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1532 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1534 * Do not assume that spinlocks can be initialized via memcpy:
1536 spin_lock_init(&ptr
->list_lock
);
1538 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1539 cachep
->nodelists
[nodeid
] = ptr
;
1543 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1544 * size of kmem_list3.
1546 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1550 for_each_online_node(node
) {
1551 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1552 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1554 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1559 * The memory after the last cpu cache pointer is used for the
1560 * the nodelists pointer.
1562 static void setup_nodelists_pointer(struct kmem_cache
*cachep
)
1564 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
1568 * Initialisation. Called after the page allocator have been initialised and
1569 * before smp_init().
1571 void __init
kmem_cache_init(void)
1573 struct cache_sizes
*sizes
;
1574 struct cache_names
*names
;
1577 kmem_cache
= &kmem_cache_boot
;
1578 setup_nodelists_pointer(kmem_cache
);
1580 if (num_possible_nodes() == 1)
1581 use_alien_caches
= 0;
1583 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1584 kmem_list3_init(&initkmem_list3
[i
]);
1586 set_up_list3s(kmem_cache
, CACHE_CACHE
);
1589 * Fragmentation resistance on low memory - only use bigger
1590 * page orders on machines with more than 32MB of memory if
1591 * not overridden on the command line.
1593 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1594 slab_max_order
= SLAB_MAX_ORDER_HI
;
1596 /* Bootstrap is tricky, because several objects are allocated
1597 * from caches that do not exist yet:
1598 * 1) initialize the kmem_cache cache: it contains the struct
1599 * kmem_cache structures of all caches, except kmem_cache itself:
1600 * kmem_cache is statically allocated.
1601 * Initially an __init data area is used for the head array and the
1602 * kmem_list3 structures, it's replaced with a kmalloc allocated
1603 * array at the end of the bootstrap.
1604 * 2) Create the first kmalloc cache.
1605 * The struct kmem_cache for the new cache is allocated normally.
1606 * An __init data area is used for the head array.
1607 * 3) Create the remaining kmalloc caches, with minimally sized
1609 * 4) Replace the __init data head arrays for kmem_cache and the first
1610 * kmalloc cache with kmalloc allocated arrays.
1611 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1612 * the other cache's with kmalloc allocated memory.
1613 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1616 /* 1) create the kmem_cache */
1619 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1621 create_boot_cache(kmem_cache
, "kmem_cache",
1622 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1623 nr_node_ids
* sizeof(struct kmem_list3
*),
1624 SLAB_HWCACHE_ALIGN
);
1625 list_add(&kmem_cache
->list
, &slab_caches
);
1627 /* 2+3) create the kmalloc caches */
1628 sizes
= malloc_sizes
;
1629 names
= cache_names
;
1632 * Initialize the caches that provide memory for the array cache and the
1633 * kmem_list3 structures first. Without this, further allocations will
1637 sizes
[INDEX_AC
].cs_cachep
= create_kmalloc_cache(names
[INDEX_AC
].name
,
1638 sizes
[INDEX_AC
].cs_size
, ARCH_KMALLOC_FLAGS
);
1640 if (INDEX_AC
!= INDEX_L3
)
1641 sizes
[INDEX_L3
].cs_cachep
=
1642 create_kmalloc_cache(names
[INDEX_L3
].name
,
1643 sizes
[INDEX_L3
].cs_size
, ARCH_KMALLOC_FLAGS
);
1645 slab_early_init
= 0;
1647 while (sizes
->cs_size
!= ULONG_MAX
) {
1649 * For performance, all the general caches are L1 aligned.
1650 * This should be particularly beneficial on SMP boxes, as it
1651 * eliminates "false sharing".
1652 * Note for systems short on memory removing the alignment will
1653 * allow tighter packing of the smaller caches.
1655 if (!sizes
->cs_cachep
)
1656 sizes
->cs_cachep
= create_kmalloc_cache(names
->name
,
1657 sizes
->cs_size
, ARCH_KMALLOC_FLAGS
);
1659 #ifdef CONFIG_ZONE_DMA
1660 sizes
->cs_dmacachep
= create_kmalloc_cache(
1661 names
->name_dma
, sizes
->cs_size
,
1662 SLAB_CACHE_DMA
|ARCH_KMALLOC_FLAGS
);
1667 /* 4) Replace the bootstrap head arrays */
1669 struct array_cache
*ptr
;
1671 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1673 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1674 sizeof(struct arraycache_init
));
1676 * Do not assume that spinlocks can be initialized via memcpy:
1678 spin_lock_init(&ptr
->lock
);
1680 kmem_cache
->array
[smp_processor_id()] = ptr
;
1682 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1684 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1685 != &initarray_generic
.cache
);
1686 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1687 sizeof(struct arraycache_init
));
1689 * Do not assume that spinlocks can be initialized via memcpy:
1691 spin_lock_init(&ptr
->lock
);
1693 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1696 /* 5) Replace the bootstrap kmem_list3's */
1700 for_each_online_node(nid
) {
1701 init_list(kmem_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1703 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1704 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1706 if (INDEX_AC
!= INDEX_L3
) {
1707 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1708 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1716 void __init
kmem_cache_init_late(void)
1718 struct kmem_cache
*cachep
;
1722 /* 6) resize the head arrays to their final sizes */
1723 mutex_lock(&slab_mutex
);
1724 list_for_each_entry(cachep
, &slab_caches
, list
)
1725 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1727 mutex_unlock(&slab_mutex
);
1729 /* Annotate slab for lockdep -- annotate the malloc caches */
1736 * Register a cpu startup notifier callback that initializes
1737 * cpu_cache_get for all new cpus
1739 register_cpu_notifier(&cpucache_notifier
);
1743 * Register a memory hotplug callback that initializes and frees
1746 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1750 * The reap timers are started later, with a module init call: That part
1751 * of the kernel is not yet operational.
1755 static int __init
cpucache_init(void)
1760 * Register the timers that return unneeded pages to the page allocator
1762 for_each_online_cpu(cpu
)
1763 start_cpu_timer(cpu
);
1769 __initcall(cpucache_init
);
1771 static noinline
void
1772 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1774 struct kmem_list3
*l3
;
1776 unsigned long flags
;
1780 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1782 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1783 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1785 for_each_online_node(node
) {
1786 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1787 unsigned long active_slabs
= 0, num_slabs
= 0;
1789 l3
= cachep
->nodelists
[node
];
1793 spin_lock_irqsave(&l3
->list_lock
, flags
);
1794 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
1795 active_objs
+= cachep
->num
;
1798 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
1799 active_objs
+= slabp
->inuse
;
1802 list_for_each_entry(slabp
, &l3
->slabs_free
, list
)
1805 free_objects
+= l3
->free_objects
;
1806 spin_unlock_irqrestore(&l3
->list_lock
, flags
);
1808 num_slabs
+= active_slabs
;
1809 num_objs
= num_slabs
* cachep
->num
;
1811 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1812 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1818 * Interface to system's page allocator. No need to hold the cache-lock.
1820 * If we requested dmaable memory, we will get it. Even if we
1821 * did not request dmaable memory, we might get it, but that
1822 * would be relatively rare and ignorable.
1824 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1832 * Nommu uses slab's for process anonymous memory allocations, and thus
1833 * requires __GFP_COMP to properly refcount higher order allocations
1835 flags
|= __GFP_COMP
;
1838 flags
|= cachep
->allocflags
;
1839 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1840 flags
|= __GFP_RECLAIMABLE
;
1842 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1844 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1845 slab_out_of_memory(cachep
, flags
, nodeid
);
1849 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1850 if (unlikely(page
->pfmemalloc
))
1851 pfmemalloc_active
= true;
1853 nr_pages
= (1 << cachep
->gfporder
);
1854 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1855 add_zone_page_state(page_zone(page
),
1856 NR_SLAB_RECLAIMABLE
, nr_pages
);
1858 add_zone_page_state(page_zone(page
),
1859 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1860 for (i
= 0; i
< nr_pages
; i
++) {
1861 __SetPageSlab(page
+ i
);
1863 if (page
->pfmemalloc
)
1864 SetPageSlabPfmemalloc(page
+ i
);
1867 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1868 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1871 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1873 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1876 return page_address(page
);
1880 * Interface to system's page release.
1882 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1884 unsigned long i
= (1 << cachep
->gfporder
);
1885 struct page
*page
= virt_to_page(addr
);
1886 const unsigned long nr_freed
= i
;
1888 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1890 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1891 sub_zone_page_state(page_zone(page
),
1892 NR_SLAB_RECLAIMABLE
, nr_freed
);
1894 sub_zone_page_state(page_zone(page
),
1895 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1897 BUG_ON(!PageSlab(page
));
1898 __ClearPageSlabPfmemalloc(page
);
1899 __ClearPageSlab(page
);
1902 if (current
->reclaim_state
)
1903 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1904 free_pages((unsigned long)addr
, cachep
->gfporder
);
1907 static void kmem_rcu_free(struct rcu_head
*head
)
1909 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1910 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1912 kmem_freepages(cachep
, slab_rcu
->addr
);
1913 if (OFF_SLAB(cachep
))
1914 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1919 #ifdef CONFIG_DEBUG_PAGEALLOC
1920 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1921 unsigned long caller
)
1923 int size
= cachep
->object_size
;
1925 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1927 if (size
< 5 * sizeof(unsigned long))
1930 *addr
++ = 0x12345678;
1932 *addr
++ = smp_processor_id();
1933 size
-= 3 * sizeof(unsigned long);
1935 unsigned long *sptr
= &caller
;
1936 unsigned long svalue
;
1938 while (!kstack_end(sptr
)) {
1940 if (kernel_text_address(svalue
)) {
1942 size
-= sizeof(unsigned long);
1943 if (size
<= sizeof(unsigned long))
1949 *addr
++ = 0x87654321;
1953 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1955 int size
= cachep
->object_size
;
1956 addr
= &((char *)addr
)[obj_offset(cachep
)];
1958 memset(addr
, val
, size
);
1959 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1962 static void dump_line(char *data
, int offset
, int limit
)
1965 unsigned char error
= 0;
1968 printk(KERN_ERR
"%03x: ", offset
);
1969 for (i
= 0; i
< limit
; i
++) {
1970 if (data
[offset
+ i
] != POISON_FREE
) {
1971 error
= data
[offset
+ i
];
1975 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1976 &data
[offset
], limit
, 1);
1978 if (bad_count
== 1) {
1979 error
^= POISON_FREE
;
1980 if (!(error
& (error
- 1))) {
1981 printk(KERN_ERR
"Single bit error detected. Probably "
1984 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1987 printk(KERN_ERR
"Run a memory test tool.\n");
1996 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
2001 if (cachep
->flags
& SLAB_RED_ZONE
) {
2002 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
2003 *dbg_redzone1(cachep
, objp
),
2004 *dbg_redzone2(cachep
, objp
));
2007 if (cachep
->flags
& SLAB_STORE_USER
) {
2008 printk(KERN_ERR
"Last user: [<%p>]",
2009 *dbg_userword(cachep
, objp
));
2010 print_symbol("(%s)",
2011 (unsigned long)*dbg_userword(cachep
, objp
));
2014 realobj
= (char *)objp
+ obj_offset(cachep
);
2015 size
= cachep
->object_size
;
2016 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
2019 if (i
+ limit
> size
)
2021 dump_line(realobj
, i
, limit
);
2025 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
2031 realobj
= (char *)objp
+ obj_offset(cachep
);
2032 size
= cachep
->object_size
;
2034 for (i
= 0; i
< size
; i
++) {
2035 char exp
= POISON_FREE
;
2038 if (realobj
[i
] != exp
) {
2044 "Slab corruption (%s): %s start=%p, len=%d\n",
2045 print_tainted(), cachep
->name
, realobj
, size
);
2046 print_objinfo(cachep
, objp
, 0);
2048 /* Hexdump the affected line */
2051 if (i
+ limit
> size
)
2053 dump_line(realobj
, i
, limit
);
2056 /* Limit to 5 lines */
2062 /* Print some data about the neighboring objects, if they
2065 struct slab
*slabp
= virt_to_slab(objp
);
2068 objnr
= obj_to_index(cachep
, slabp
, objp
);
2070 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
2071 realobj
= (char *)objp
+ obj_offset(cachep
);
2072 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
2074 print_objinfo(cachep
, objp
, 2);
2076 if (objnr
+ 1 < cachep
->num
) {
2077 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
2078 realobj
= (char *)objp
+ obj_offset(cachep
);
2079 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
2081 print_objinfo(cachep
, objp
, 2);
2088 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2091 for (i
= 0; i
< cachep
->num
; i
++) {
2092 void *objp
= index_to_obj(cachep
, slabp
, i
);
2094 if (cachep
->flags
& SLAB_POISON
) {
2095 #ifdef CONFIG_DEBUG_PAGEALLOC
2096 if (cachep
->size
% PAGE_SIZE
== 0 &&
2098 kernel_map_pages(virt_to_page(objp
),
2099 cachep
->size
/ PAGE_SIZE
, 1);
2101 check_poison_obj(cachep
, objp
);
2103 check_poison_obj(cachep
, objp
);
2106 if (cachep
->flags
& SLAB_RED_ZONE
) {
2107 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2108 slab_error(cachep
, "start of a freed object "
2110 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2111 slab_error(cachep
, "end of a freed object "
2117 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2123 * slab_destroy - destroy and release all objects in a slab
2124 * @cachep: cache pointer being destroyed
2125 * @slabp: slab pointer being destroyed
2127 * Destroy all the objs in a slab, and release the mem back to the system.
2128 * Before calling the slab must have been unlinked from the cache. The
2129 * cache-lock is not held/needed.
2131 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2133 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2135 slab_destroy_debugcheck(cachep
, slabp
);
2136 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2137 struct slab_rcu
*slab_rcu
;
2139 slab_rcu
= (struct slab_rcu
*)slabp
;
2140 slab_rcu
->cachep
= cachep
;
2141 slab_rcu
->addr
= addr
;
2142 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2144 kmem_freepages(cachep
, addr
);
2145 if (OFF_SLAB(cachep
))
2146 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2151 * calculate_slab_order - calculate size (page order) of slabs
2152 * @cachep: pointer to the cache that is being created
2153 * @size: size of objects to be created in this cache.
2154 * @align: required alignment for the objects.
2155 * @flags: slab allocation flags
2157 * Also calculates the number of objects per slab.
2159 * This could be made much more intelligent. For now, try to avoid using
2160 * high order pages for slabs. When the gfp() functions are more friendly
2161 * towards high-order requests, this should be changed.
2163 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2164 size_t size
, size_t align
, unsigned long flags
)
2166 unsigned long offslab_limit
;
2167 size_t left_over
= 0;
2170 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2174 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2178 if (flags
& CFLGS_OFF_SLAB
) {
2180 * Max number of objs-per-slab for caches which
2181 * use off-slab slabs. Needed to avoid a possible
2182 * looping condition in cache_grow().
2184 offslab_limit
= size
- sizeof(struct slab
);
2185 offslab_limit
/= sizeof(kmem_bufctl_t
);
2187 if (num
> offslab_limit
)
2191 /* Found something acceptable - save it away */
2193 cachep
->gfporder
= gfporder
;
2194 left_over
= remainder
;
2197 * A VFS-reclaimable slab tends to have most allocations
2198 * as GFP_NOFS and we really don't want to have to be allocating
2199 * higher-order pages when we are unable to shrink dcache.
2201 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2205 * Large number of objects is good, but very large slabs are
2206 * currently bad for the gfp()s.
2208 if (gfporder
>= slab_max_order
)
2212 * Acceptable internal fragmentation?
2214 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2220 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2222 if (slab_state
>= FULL
)
2223 return enable_cpucache(cachep
, gfp
);
2225 if (slab_state
== DOWN
) {
2227 * Note: Creation of first cache (kmem_cache).
2228 * The setup_list3s is taken care
2229 * of by the caller of __kmem_cache_create
2231 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2232 slab_state
= PARTIAL
;
2233 } else if (slab_state
== PARTIAL
) {
2235 * Note: the second kmem_cache_create must create the cache
2236 * that's used by kmalloc(24), otherwise the creation of
2237 * further caches will BUG().
2239 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2242 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2243 * the second cache, then we need to set up all its list3s,
2244 * otherwise the creation of further caches will BUG().
2246 set_up_list3s(cachep
, SIZE_AC
);
2247 if (INDEX_AC
== INDEX_L3
)
2248 slab_state
= PARTIAL_L3
;
2250 slab_state
= PARTIAL_ARRAYCACHE
;
2252 /* Remaining boot caches */
2253 cachep
->array
[smp_processor_id()] =
2254 kmalloc(sizeof(struct arraycache_init
), gfp
);
2256 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2257 set_up_list3s(cachep
, SIZE_L3
);
2258 slab_state
= PARTIAL_L3
;
2261 for_each_online_node(node
) {
2262 cachep
->nodelists
[node
] =
2263 kmalloc_node(sizeof(struct kmem_list3
),
2265 BUG_ON(!cachep
->nodelists
[node
]);
2266 kmem_list3_init(cachep
->nodelists
[node
]);
2270 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2271 jiffies
+ REAPTIMEOUT_LIST3
+
2272 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2274 cpu_cache_get(cachep
)->avail
= 0;
2275 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2276 cpu_cache_get(cachep
)->batchcount
= 1;
2277 cpu_cache_get(cachep
)->touched
= 0;
2278 cachep
->batchcount
= 1;
2279 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2284 * __kmem_cache_create - Create a cache.
2285 * @cachep: cache management descriptor
2286 * @flags: SLAB flags
2288 * Returns a ptr to the cache on success, NULL on failure.
2289 * Cannot be called within a int, but can be interrupted.
2290 * The @ctor is run when new pages are allocated by the cache.
2294 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2295 * to catch references to uninitialised memory.
2297 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2298 * for buffer overruns.
2300 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2301 * cacheline. This can be beneficial if you're counting cycles as closely
2305 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2307 size_t left_over
, slab_size
, ralign
;
2310 size_t size
= cachep
->size
;
2315 * Enable redzoning and last user accounting, except for caches with
2316 * large objects, if the increased size would increase the object size
2317 * above the next power of two: caches with object sizes just above a
2318 * power of two have a significant amount of internal fragmentation.
2320 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2321 2 * sizeof(unsigned long long)))
2322 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2323 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2324 flags
|= SLAB_POISON
;
2326 if (flags
& SLAB_DESTROY_BY_RCU
)
2327 BUG_ON(flags
& SLAB_POISON
);
2331 * Check that size is in terms of words. This is needed to avoid
2332 * unaligned accesses for some archs when redzoning is used, and makes
2333 * sure any on-slab bufctl's are also correctly aligned.
2335 if (size
& (BYTES_PER_WORD
- 1)) {
2336 size
+= (BYTES_PER_WORD
- 1);
2337 size
&= ~(BYTES_PER_WORD
- 1);
2341 * Redzoning and user store require word alignment or possibly larger.
2342 * Note this will be overridden by architecture or caller mandated
2343 * alignment if either is greater than BYTES_PER_WORD.
2345 if (flags
& SLAB_STORE_USER
)
2346 ralign
= BYTES_PER_WORD
;
2348 if (flags
& SLAB_RED_ZONE
) {
2349 ralign
= REDZONE_ALIGN
;
2350 /* If redzoning, ensure that the second redzone is suitably
2351 * aligned, by adjusting the object size accordingly. */
2352 size
+= REDZONE_ALIGN
- 1;
2353 size
&= ~(REDZONE_ALIGN
- 1);
2356 /* 3) caller mandated alignment */
2357 if (ralign
< cachep
->align
) {
2358 ralign
= cachep
->align
;
2360 /* disable debug if necessary */
2361 if (ralign
> __alignof__(unsigned long long))
2362 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2366 cachep
->align
= ralign
;
2368 if (slab_is_available())
2373 setup_nodelists_pointer(cachep
);
2377 * Both debugging options require word-alignment which is calculated
2380 if (flags
& SLAB_RED_ZONE
) {
2381 /* add space for red zone words */
2382 cachep
->obj_offset
+= sizeof(unsigned long long);
2383 size
+= 2 * sizeof(unsigned long long);
2385 if (flags
& SLAB_STORE_USER
) {
2386 /* user store requires one word storage behind the end of
2387 * the real object. But if the second red zone needs to be
2388 * aligned to 64 bits, we must allow that much space.
2390 if (flags
& SLAB_RED_ZONE
)
2391 size
+= REDZONE_ALIGN
;
2393 size
+= BYTES_PER_WORD
;
2395 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2396 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2397 && cachep
->object_size
> cache_line_size()
2398 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2399 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2406 * Determine if the slab management is 'on' or 'off' slab.
2407 * (bootstrapping cannot cope with offslab caches so don't do
2408 * it too early on. Always use on-slab management when
2409 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2411 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2412 !(flags
& SLAB_NOLEAKTRACE
))
2414 * Size is large, assume best to place the slab management obj
2415 * off-slab (should allow better packing of objs).
2417 flags
|= CFLGS_OFF_SLAB
;
2419 size
= ALIGN(size
, cachep
->align
);
2421 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2426 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2427 + sizeof(struct slab
), cachep
->align
);
2430 * If the slab has been placed off-slab, and we have enough space then
2431 * move it on-slab. This is at the expense of any extra colouring.
2433 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2434 flags
&= ~CFLGS_OFF_SLAB
;
2435 left_over
-= slab_size
;
2438 if (flags
& CFLGS_OFF_SLAB
) {
2439 /* really off slab. No need for manual alignment */
2441 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2443 #ifdef CONFIG_PAGE_POISONING
2444 /* If we're going to use the generic kernel_map_pages()
2445 * poisoning, then it's going to smash the contents of
2446 * the redzone and userword anyhow, so switch them off.
2448 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2449 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2453 cachep
->colour_off
= cache_line_size();
2454 /* Offset must be a multiple of the alignment. */
2455 if (cachep
->colour_off
< cachep
->align
)
2456 cachep
->colour_off
= cachep
->align
;
2457 cachep
->colour
= left_over
/ cachep
->colour_off
;
2458 cachep
->slab_size
= slab_size
;
2459 cachep
->flags
= flags
;
2460 cachep
->allocflags
= 0;
2461 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2462 cachep
->allocflags
|= GFP_DMA
;
2463 cachep
->size
= size
;
2464 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2466 if (flags
& CFLGS_OFF_SLAB
) {
2467 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2469 * This is a possibility for one of the malloc_sizes caches.
2470 * But since we go off slab only for object size greater than
2471 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2472 * this should not happen at all.
2473 * But leave a BUG_ON for some lucky dude.
2475 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2478 err
= setup_cpu_cache(cachep
, gfp
);
2480 __kmem_cache_shutdown(cachep
);
2484 if (flags
& SLAB_DEBUG_OBJECTS
) {
2486 * Would deadlock through slab_destroy()->call_rcu()->
2487 * debug_object_activate()->kmem_cache_alloc().
2489 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2491 slab_set_debugobj_lock_classes(cachep
);
2498 static void check_irq_off(void)
2500 BUG_ON(!irqs_disabled());
2503 static void check_irq_on(void)
2505 BUG_ON(irqs_disabled());
2508 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2512 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2516 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2520 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2525 #define check_irq_off() do { } while(0)
2526 #define check_irq_on() do { } while(0)
2527 #define check_spinlock_acquired(x) do { } while(0)
2528 #define check_spinlock_acquired_node(x, y) do { } while(0)
2531 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2532 struct array_cache
*ac
,
2533 int force
, int node
);
2535 static void do_drain(void *arg
)
2537 struct kmem_cache
*cachep
= arg
;
2538 struct array_cache
*ac
;
2539 int node
= numa_mem_id();
2542 ac
= cpu_cache_get(cachep
);
2543 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2544 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2545 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2549 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2551 struct kmem_list3
*l3
;
2554 on_each_cpu(do_drain
, cachep
, 1);
2556 for_each_online_node(node
) {
2557 l3
= cachep
->nodelists
[node
];
2558 if (l3
&& l3
->alien
)
2559 drain_alien_cache(cachep
, l3
->alien
);
2562 for_each_online_node(node
) {
2563 l3
= cachep
->nodelists
[node
];
2565 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2570 * Remove slabs from the list of free slabs.
2571 * Specify the number of slabs to drain in tofree.
2573 * Returns the actual number of slabs released.
2575 static int drain_freelist(struct kmem_cache
*cache
,
2576 struct kmem_list3
*l3
, int tofree
)
2578 struct list_head
*p
;
2583 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2585 spin_lock_irq(&l3
->list_lock
);
2586 p
= l3
->slabs_free
.prev
;
2587 if (p
== &l3
->slabs_free
) {
2588 spin_unlock_irq(&l3
->list_lock
);
2592 slabp
= list_entry(p
, struct slab
, list
);
2594 BUG_ON(slabp
->inuse
);
2596 list_del(&slabp
->list
);
2598 * Safe to drop the lock. The slab is no longer linked
2601 l3
->free_objects
-= cache
->num
;
2602 spin_unlock_irq(&l3
->list_lock
);
2603 slab_destroy(cache
, slabp
);
2610 /* Called with slab_mutex held to protect against cpu hotplug */
2611 static int __cache_shrink(struct kmem_cache
*cachep
)
2614 struct kmem_list3
*l3
;
2616 drain_cpu_caches(cachep
);
2619 for_each_online_node(i
) {
2620 l3
= cachep
->nodelists
[i
];
2624 drain_freelist(cachep
, l3
, l3
->free_objects
);
2626 ret
+= !list_empty(&l3
->slabs_full
) ||
2627 !list_empty(&l3
->slabs_partial
);
2629 return (ret
? 1 : 0);
2633 * kmem_cache_shrink - Shrink a cache.
2634 * @cachep: The cache to shrink.
2636 * Releases as many slabs as possible for a cache.
2637 * To help debugging, a zero exit status indicates all slabs were released.
2639 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2642 BUG_ON(!cachep
|| in_interrupt());
2645 mutex_lock(&slab_mutex
);
2646 ret
= __cache_shrink(cachep
);
2647 mutex_unlock(&slab_mutex
);
2651 EXPORT_SYMBOL(kmem_cache_shrink
);
2653 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2656 struct kmem_list3
*l3
;
2657 int rc
= __cache_shrink(cachep
);
2662 for_each_online_cpu(i
)
2663 kfree(cachep
->array
[i
]);
2665 /* NUMA: free the list3 structures */
2666 for_each_online_node(i
) {
2667 l3
= cachep
->nodelists
[i
];
2670 free_alien_cache(l3
->alien
);
2678 * Get the memory for a slab management obj.
2679 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2680 * always come from malloc_sizes caches. The slab descriptor cannot
2681 * come from the same cache which is getting created because,
2682 * when we are searching for an appropriate cache for these
2683 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2684 * If we are creating a malloc_sizes cache here it would not be visible to
2685 * kmem_find_general_cachep till the initialization is complete.
2686 * Hence we cannot have slabp_cache same as the original cache.
2688 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2689 int colour_off
, gfp_t local_flags
,
2694 if (OFF_SLAB(cachep
)) {
2695 /* Slab management obj is off-slab. */
2696 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2697 local_flags
, nodeid
);
2699 * If the first object in the slab is leaked (it's allocated
2700 * but no one has a reference to it), we want to make sure
2701 * kmemleak does not treat the ->s_mem pointer as a reference
2702 * to the object. Otherwise we will not report the leak.
2704 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2709 slabp
= objp
+ colour_off
;
2710 colour_off
+= cachep
->slab_size
;
2713 slabp
->colouroff
= colour_off
;
2714 slabp
->s_mem
= objp
+ colour_off
;
2715 slabp
->nodeid
= nodeid
;
2720 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2722 return (kmem_bufctl_t
*) (slabp
+ 1);
2725 static void cache_init_objs(struct kmem_cache
*cachep
,
2730 for (i
= 0; i
< cachep
->num
; i
++) {
2731 void *objp
= index_to_obj(cachep
, slabp
, i
);
2733 /* need to poison the objs? */
2734 if (cachep
->flags
& SLAB_POISON
)
2735 poison_obj(cachep
, objp
, POISON_FREE
);
2736 if (cachep
->flags
& SLAB_STORE_USER
)
2737 *dbg_userword(cachep
, objp
) = NULL
;
2739 if (cachep
->flags
& SLAB_RED_ZONE
) {
2740 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2741 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2744 * Constructors are not allowed to allocate memory from the same
2745 * cache which they are a constructor for. Otherwise, deadlock.
2746 * They must also be threaded.
2748 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2749 cachep
->ctor(objp
+ obj_offset(cachep
));
2751 if (cachep
->flags
& SLAB_RED_ZONE
) {
2752 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2753 slab_error(cachep
, "constructor overwrote the"
2754 " end of an object");
2755 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2756 slab_error(cachep
, "constructor overwrote the"
2757 " start of an object");
2759 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2760 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2761 kernel_map_pages(virt_to_page(objp
),
2762 cachep
->size
/ PAGE_SIZE
, 0);
2767 slab_bufctl(slabp
)[i
] = i
+ 1;
2769 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2772 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2774 if (CONFIG_ZONE_DMA_FLAG
) {
2775 if (flags
& GFP_DMA
)
2776 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2778 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2782 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2785 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2789 next
= slab_bufctl(slabp
)[slabp
->free
];
2791 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2792 WARN_ON(slabp
->nodeid
!= nodeid
);
2799 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2800 void *objp
, int nodeid
)
2802 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2805 /* Verify that the slab belongs to the intended node */
2806 WARN_ON(slabp
->nodeid
!= nodeid
);
2808 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2809 printk(KERN_ERR
"slab: double free detected in cache "
2810 "'%s', objp %p\n", cachep
->name
, objp
);
2814 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2815 slabp
->free
= objnr
;
2820 * Map pages beginning at addr to the given cache and slab. This is required
2821 * for the slab allocator to be able to lookup the cache and slab of a
2822 * virtual address for kfree, ksize, and slab debugging.
2824 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2830 page
= virt_to_page(addr
);
2833 if (likely(!PageCompound(page
)))
2834 nr_pages
<<= cache
->gfporder
;
2837 page
->slab_cache
= cache
;
2838 page
->slab_page
= slab
;
2840 } while (--nr_pages
);
2844 * Grow (by 1) the number of slabs within a cache. This is called by
2845 * kmem_cache_alloc() when there are no active objs left in a cache.
2847 static int cache_grow(struct kmem_cache
*cachep
,
2848 gfp_t flags
, int nodeid
, void *objp
)
2853 struct kmem_list3
*l3
;
2856 * Be lazy and only check for valid flags here, keeping it out of the
2857 * critical path in kmem_cache_alloc().
2859 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2860 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2862 /* Take the l3 list lock to change the colour_next on this node */
2864 l3
= cachep
->nodelists
[nodeid
];
2865 spin_lock(&l3
->list_lock
);
2867 /* Get colour for the slab, and cal the next value. */
2868 offset
= l3
->colour_next
;
2870 if (l3
->colour_next
>= cachep
->colour
)
2871 l3
->colour_next
= 0;
2872 spin_unlock(&l3
->list_lock
);
2874 offset
*= cachep
->colour_off
;
2876 if (local_flags
& __GFP_WAIT
)
2880 * The test for missing atomic flag is performed here, rather than
2881 * the more obvious place, simply to reduce the critical path length
2882 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2883 * will eventually be caught here (where it matters).
2885 kmem_flagcheck(cachep
, flags
);
2888 * Get mem for the objs. Attempt to allocate a physical page from
2892 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2896 /* Get slab management. */
2897 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2898 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2902 slab_map_pages(cachep
, slabp
, objp
);
2904 cache_init_objs(cachep
, slabp
);
2906 if (local_flags
& __GFP_WAIT
)
2907 local_irq_disable();
2909 spin_lock(&l3
->list_lock
);
2911 /* Make slab active. */
2912 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2913 STATS_INC_GROWN(cachep
);
2914 l3
->free_objects
+= cachep
->num
;
2915 spin_unlock(&l3
->list_lock
);
2918 kmem_freepages(cachep
, objp
);
2920 if (local_flags
& __GFP_WAIT
)
2921 local_irq_disable();
2928 * Perform extra freeing checks:
2929 * - detect bad pointers.
2930 * - POISON/RED_ZONE checking
2932 static void kfree_debugcheck(const void *objp
)
2934 if (!virt_addr_valid(objp
)) {
2935 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2936 (unsigned long)objp
);
2941 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2943 unsigned long long redzone1
, redzone2
;
2945 redzone1
= *dbg_redzone1(cache
, obj
);
2946 redzone2
= *dbg_redzone2(cache
, obj
);
2951 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2954 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2955 slab_error(cache
, "double free detected");
2957 slab_error(cache
, "memory outside object was overwritten");
2959 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2960 obj
, redzone1
, redzone2
);
2963 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2964 unsigned long caller
)
2970 BUG_ON(virt_to_cache(objp
) != cachep
);
2972 objp
-= obj_offset(cachep
);
2973 kfree_debugcheck(objp
);
2974 page
= virt_to_head_page(objp
);
2976 slabp
= page
->slab_page
;
2978 if (cachep
->flags
& SLAB_RED_ZONE
) {
2979 verify_redzone_free(cachep
, objp
);
2980 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2981 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2983 if (cachep
->flags
& SLAB_STORE_USER
)
2984 *dbg_userword(cachep
, objp
) = (void *)caller
;
2986 objnr
= obj_to_index(cachep
, slabp
, objp
);
2988 BUG_ON(objnr
>= cachep
->num
);
2989 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2991 #ifdef CONFIG_DEBUG_SLAB_LEAK
2992 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2994 if (cachep
->flags
& SLAB_POISON
) {
2995 #ifdef CONFIG_DEBUG_PAGEALLOC
2996 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2997 store_stackinfo(cachep
, objp
, caller
);
2998 kernel_map_pages(virt_to_page(objp
),
2999 cachep
->size
/ PAGE_SIZE
, 0);
3001 poison_obj(cachep
, objp
, POISON_FREE
);
3004 poison_obj(cachep
, objp
, POISON_FREE
);
3010 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
3015 /* Check slab's freelist to see if this obj is there. */
3016 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
3018 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3021 if (entries
!= cachep
->num
- slabp
->inuse
) {
3023 printk(KERN_ERR
"slab: Internal list corruption detected in "
3024 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3025 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
3027 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3028 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3034 #define kfree_debugcheck(x) do { } while(0)
3035 #define cache_free_debugcheck(x,objp,z) (objp)
3036 #define check_slabp(x,y) do { } while(0)
3039 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
3043 struct kmem_list3
*l3
;
3044 struct array_cache
*ac
;
3048 node
= numa_mem_id();
3049 if (unlikely(force_refill
))
3052 ac
= cpu_cache_get(cachep
);
3053 batchcount
= ac
->batchcount
;
3054 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3056 * If there was little recent activity on this cache, then
3057 * perform only a partial refill. Otherwise we could generate
3060 batchcount
= BATCHREFILL_LIMIT
;
3062 l3
= cachep
->nodelists
[node
];
3064 BUG_ON(ac
->avail
> 0 || !l3
);
3065 spin_lock(&l3
->list_lock
);
3067 /* See if we can refill from the shared array */
3068 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3069 l3
->shared
->touched
= 1;
3073 while (batchcount
> 0) {
3074 struct list_head
*entry
;
3076 /* Get slab alloc is to come from. */
3077 entry
= l3
->slabs_partial
.next
;
3078 if (entry
== &l3
->slabs_partial
) {
3079 l3
->free_touched
= 1;
3080 entry
= l3
->slabs_free
.next
;
3081 if (entry
== &l3
->slabs_free
)
3085 slabp
= list_entry(entry
, struct slab
, list
);
3086 check_slabp(cachep
, slabp
);
3087 check_spinlock_acquired(cachep
);
3090 * The slab was either on partial or free list so
3091 * there must be at least one object available for
3094 BUG_ON(slabp
->inuse
>= cachep
->num
);
3096 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3097 STATS_INC_ALLOCED(cachep
);
3098 STATS_INC_ACTIVE(cachep
);
3099 STATS_SET_HIGH(cachep
);
3101 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
3104 check_slabp(cachep
, slabp
);
3106 /* move slabp to correct slabp list: */
3107 list_del(&slabp
->list
);
3108 if (slabp
->free
== BUFCTL_END
)
3109 list_add(&slabp
->list
, &l3
->slabs_full
);
3111 list_add(&slabp
->list
, &l3
->slabs_partial
);
3115 l3
->free_objects
-= ac
->avail
;
3117 spin_unlock(&l3
->list_lock
);
3119 if (unlikely(!ac
->avail
)) {
3122 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3124 /* cache_grow can reenable interrupts, then ac could change. */
3125 ac
= cpu_cache_get(cachep
);
3126 node
= numa_mem_id();
3128 /* no objects in sight? abort */
3129 if (!x
&& (ac
->avail
== 0 || force_refill
))
3132 if (!ac
->avail
) /* objects refilled by interrupt? */
3137 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
3140 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3143 might_sleep_if(flags
& __GFP_WAIT
);
3145 kmem_flagcheck(cachep
, flags
);
3150 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3151 gfp_t flags
, void *objp
, unsigned long caller
)
3155 if (cachep
->flags
& SLAB_POISON
) {
3156 #ifdef CONFIG_DEBUG_PAGEALLOC
3157 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3158 kernel_map_pages(virt_to_page(objp
),
3159 cachep
->size
/ PAGE_SIZE
, 1);
3161 check_poison_obj(cachep
, objp
);
3163 check_poison_obj(cachep
, objp
);
3165 poison_obj(cachep
, objp
, POISON_INUSE
);
3167 if (cachep
->flags
& SLAB_STORE_USER
)
3168 *dbg_userword(cachep
, objp
) = (void *)caller
;
3170 if (cachep
->flags
& SLAB_RED_ZONE
) {
3171 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3172 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3173 slab_error(cachep
, "double free, or memory outside"
3174 " object was overwritten");
3176 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3177 objp
, *dbg_redzone1(cachep
, objp
),
3178 *dbg_redzone2(cachep
, objp
));
3180 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3181 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3183 #ifdef CONFIG_DEBUG_SLAB_LEAK
3188 slabp
= virt_to_head_page(objp
)->slab_page
;
3189 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3190 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3193 objp
+= obj_offset(cachep
);
3194 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3196 if (ARCH_SLAB_MINALIGN
&&
3197 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3198 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3199 objp
, (int)ARCH_SLAB_MINALIGN
);
3204 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3207 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3209 if (cachep
== kmem_cache
)
3212 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3215 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3218 struct array_cache
*ac
;
3219 bool force_refill
= false;
3223 ac
= cpu_cache_get(cachep
);
3224 if (likely(ac
->avail
)) {
3226 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3229 * Allow for the possibility all avail objects are not allowed
3230 * by the current flags
3233 STATS_INC_ALLOCHIT(cachep
);
3236 force_refill
= true;
3239 STATS_INC_ALLOCMISS(cachep
);
3240 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3242 * the 'ac' may be updated by cache_alloc_refill(),
3243 * and kmemleak_erase() requires its correct value.
3245 ac
= cpu_cache_get(cachep
);
3249 * To avoid a false negative, if an object that is in one of the
3250 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3251 * treat the array pointers as a reference to the object.
3254 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3260 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3262 * If we are in_interrupt, then process context, including cpusets and
3263 * mempolicy, may not apply and should not be used for allocation policy.
3265 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3267 int nid_alloc
, nid_here
;
3269 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3271 nid_alloc
= nid_here
= numa_mem_id();
3272 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3273 nid_alloc
= cpuset_slab_spread_node();
3274 else if (current
->mempolicy
)
3275 nid_alloc
= slab_node();
3276 if (nid_alloc
!= nid_here
)
3277 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3282 * Fallback function if there was no memory available and no objects on a
3283 * certain node and fall back is permitted. First we scan all the
3284 * available nodelists for available objects. If that fails then we
3285 * perform an allocation without specifying a node. This allows the page
3286 * allocator to do its reclaim / fallback magic. We then insert the
3287 * slab into the proper nodelist and then allocate from it.
3289 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3291 struct zonelist
*zonelist
;
3295 enum zone_type high_zoneidx
= gfp_zone(flags
);
3298 unsigned int cpuset_mems_cookie
;
3300 if (flags
& __GFP_THISNODE
)
3303 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3306 cpuset_mems_cookie
= get_mems_allowed();
3307 zonelist
= node_zonelist(slab_node(), flags
);
3311 * Look through allowed nodes for objects available
3312 * from existing per node queues.
3314 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3315 nid
= zone_to_nid(zone
);
3317 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3318 cache
->nodelists
[nid
] &&
3319 cache
->nodelists
[nid
]->free_objects
) {
3320 obj
= ____cache_alloc_node(cache
,
3321 flags
| GFP_THISNODE
, nid
);
3329 * This allocation will be performed within the constraints
3330 * of the current cpuset / memory policy requirements.
3331 * We may trigger various forms of reclaim on the allowed
3332 * set and go into memory reserves if necessary.
3334 if (local_flags
& __GFP_WAIT
)
3336 kmem_flagcheck(cache
, flags
);
3337 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3338 if (local_flags
& __GFP_WAIT
)
3339 local_irq_disable();
3342 * Insert into the appropriate per node queues
3344 nid
= page_to_nid(virt_to_page(obj
));
3345 if (cache_grow(cache
, flags
, nid
, obj
)) {
3346 obj
= ____cache_alloc_node(cache
,
3347 flags
| GFP_THISNODE
, nid
);
3350 * Another processor may allocate the
3351 * objects in the slab since we are
3352 * not holding any locks.
3356 /* cache_grow already freed obj */
3362 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3368 * A interface to enable slab creation on nodeid
3370 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3373 struct list_head
*entry
;
3375 struct kmem_list3
*l3
;
3379 l3
= cachep
->nodelists
[nodeid
];
3384 spin_lock(&l3
->list_lock
);
3385 entry
= l3
->slabs_partial
.next
;
3386 if (entry
== &l3
->slabs_partial
) {
3387 l3
->free_touched
= 1;
3388 entry
= l3
->slabs_free
.next
;
3389 if (entry
== &l3
->slabs_free
)
3393 slabp
= list_entry(entry
, struct slab
, list
);
3394 check_spinlock_acquired_node(cachep
, nodeid
);
3395 check_slabp(cachep
, slabp
);
3397 STATS_INC_NODEALLOCS(cachep
);
3398 STATS_INC_ACTIVE(cachep
);
3399 STATS_SET_HIGH(cachep
);
3401 BUG_ON(slabp
->inuse
== cachep
->num
);
3403 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3404 check_slabp(cachep
, slabp
);
3406 /* move slabp to correct slabp list: */
3407 list_del(&slabp
->list
);
3409 if (slabp
->free
== BUFCTL_END
)
3410 list_add(&slabp
->list
, &l3
->slabs_full
);
3412 list_add(&slabp
->list
, &l3
->slabs_partial
);
3414 spin_unlock(&l3
->list_lock
);
3418 spin_unlock(&l3
->list_lock
);
3419 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3423 return fallback_alloc(cachep
, flags
);
3430 * kmem_cache_alloc_node - Allocate an object on the specified node
3431 * @cachep: The cache to allocate from.
3432 * @flags: See kmalloc().
3433 * @nodeid: node number of the target node.
3434 * @caller: return address of caller, used for debug information
3436 * Identical to kmem_cache_alloc but it will allocate memory on the given
3437 * node, which can improve the performance for cpu bound structures.
3439 * Fallback to other node is possible if __GFP_THISNODE is not set.
3441 static __always_inline
void *
3442 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3443 unsigned long caller
)
3445 unsigned long save_flags
;
3447 int slab_node
= numa_mem_id();
3449 flags
&= gfp_allowed_mask
;
3451 lockdep_trace_alloc(flags
);
3453 if (slab_should_failslab(cachep
, flags
))
3456 cache_alloc_debugcheck_before(cachep
, flags
);
3457 local_irq_save(save_flags
);
3459 if (nodeid
== NUMA_NO_NODE
)
3462 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3463 /* Node not bootstrapped yet */
3464 ptr
= fallback_alloc(cachep
, flags
);
3468 if (nodeid
== slab_node
) {
3470 * Use the locally cached objects if possible.
3471 * However ____cache_alloc does not allow fallback
3472 * to other nodes. It may fail while we still have
3473 * objects on other nodes available.
3475 ptr
= ____cache_alloc(cachep
, flags
);
3479 /* ___cache_alloc_node can fall back to other nodes */
3480 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3482 local_irq_restore(save_flags
);
3483 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3484 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3488 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3490 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3491 memset(ptr
, 0, cachep
->object_size
);
3496 static __always_inline
void *
3497 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3501 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3502 objp
= alternate_node_alloc(cache
, flags
);
3506 objp
= ____cache_alloc(cache
, flags
);
3509 * We may just have run out of memory on the local node.
3510 * ____cache_alloc_node() knows how to locate memory on other nodes
3513 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3520 static __always_inline
void *
3521 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3523 return ____cache_alloc(cachep
, flags
);
3526 #endif /* CONFIG_NUMA */
3528 static __always_inline
void *
3529 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3531 unsigned long save_flags
;
3534 flags
&= gfp_allowed_mask
;
3536 lockdep_trace_alloc(flags
);
3538 if (slab_should_failslab(cachep
, flags
))
3541 cache_alloc_debugcheck_before(cachep
, flags
);
3542 local_irq_save(save_flags
);
3543 objp
= __do_cache_alloc(cachep
, flags
);
3544 local_irq_restore(save_flags
);
3545 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3546 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3551 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3553 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3554 memset(objp
, 0, cachep
->object_size
);
3560 * Caller needs to acquire correct kmem_list's list_lock
3562 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3566 struct kmem_list3
*l3
;
3568 for (i
= 0; i
< nr_objects
; i
++) {
3572 clear_obj_pfmemalloc(&objpp
[i
]);
3575 slabp
= virt_to_slab(objp
);
3576 l3
= cachep
->nodelists
[node
];
3577 list_del(&slabp
->list
);
3578 check_spinlock_acquired_node(cachep
, node
);
3579 check_slabp(cachep
, slabp
);
3580 slab_put_obj(cachep
, slabp
, objp
, node
);
3581 STATS_DEC_ACTIVE(cachep
);
3583 check_slabp(cachep
, slabp
);
3585 /* fixup slab chains */
3586 if (slabp
->inuse
== 0) {
3587 if (l3
->free_objects
> l3
->free_limit
) {
3588 l3
->free_objects
-= cachep
->num
;
3589 /* No need to drop any previously held
3590 * lock here, even if we have a off-slab slab
3591 * descriptor it is guaranteed to come from
3592 * a different cache, refer to comments before
3595 slab_destroy(cachep
, slabp
);
3597 list_add(&slabp
->list
, &l3
->slabs_free
);
3600 /* Unconditionally move a slab to the end of the
3601 * partial list on free - maximum time for the
3602 * other objects to be freed, too.
3604 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3609 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3612 struct kmem_list3
*l3
;
3613 int node
= numa_mem_id();
3615 batchcount
= ac
->batchcount
;
3617 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3620 l3
= cachep
->nodelists
[node
];
3621 spin_lock(&l3
->list_lock
);
3623 struct array_cache
*shared_array
= l3
->shared
;
3624 int max
= shared_array
->limit
- shared_array
->avail
;
3626 if (batchcount
> max
)
3628 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3629 ac
->entry
, sizeof(void *) * batchcount
);
3630 shared_array
->avail
+= batchcount
;
3635 free_block(cachep
, ac
->entry
, batchcount
, node
);
3640 struct list_head
*p
;
3642 p
= l3
->slabs_free
.next
;
3643 while (p
!= &(l3
->slabs_free
)) {
3646 slabp
= list_entry(p
, struct slab
, list
);
3647 BUG_ON(slabp
->inuse
);
3652 STATS_SET_FREEABLE(cachep
, i
);
3655 spin_unlock(&l3
->list_lock
);
3656 ac
->avail
-= batchcount
;
3657 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3661 * Release an obj back to its cache. If the obj has a constructed state, it must
3662 * be in this state _before_ it is released. Called with disabled ints.
3664 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3665 unsigned long caller
)
3667 struct array_cache
*ac
= cpu_cache_get(cachep
);
3670 kmemleak_free_recursive(objp
, cachep
->flags
);
3671 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3673 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3676 * Skip calling cache_free_alien() when the platform is not numa.
3677 * This will avoid cache misses that happen while accessing slabp (which
3678 * is per page memory reference) to get nodeid. Instead use a global
3679 * variable to skip the call, which is mostly likely to be present in
3682 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3685 if (likely(ac
->avail
< ac
->limit
)) {
3686 STATS_INC_FREEHIT(cachep
);
3688 STATS_INC_FREEMISS(cachep
);
3689 cache_flusharray(cachep
, ac
);
3692 ac_put_obj(cachep
, ac
, objp
);
3696 * kmem_cache_alloc - Allocate an object
3697 * @cachep: The cache to allocate from.
3698 * @flags: See kmalloc().
3700 * Allocate an object from this cache. The flags are only relevant
3701 * if the cache has no available objects.
3703 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3705 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3707 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3708 cachep
->object_size
, cachep
->size
, flags
);
3712 EXPORT_SYMBOL(kmem_cache_alloc
);
3714 #ifdef CONFIG_TRACING
3716 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3720 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3722 trace_kmalloc(_RET_IP_
, ret
,
3723 size
, cachep
->size
, flags
);
3726 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3730 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3732 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3734 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3735 cachep
->object_size
, cachep
->size
,
3740 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3742 #ifdef CONFIG_TRACING
3743 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3750 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3752 trace_kmalloc_node(_RET_IP_
, ret
,
3757 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3760 static __always_inline
void *
3761 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3763 struct kmem_cache
*cachep
;
3765 cachep
= kmem_find_general_cachep(size
, flags
);
3766 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3768 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3771 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3772 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3774 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3776 EXPORT_SYMBOL(__kmalloc_node
);
3778 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3779 int node
, unsigned long caller
)
3781 return __do_kmalloc_node(size
, flags
, node
, caller
);
3783 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3785 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3787 return __do_kmalloc_node(size
, flags
, node
, 0);
3789 EXPORT_SYMBOL(__kmalloc_node
);
3790 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3791 #endif /* CONFIG_NUMA */
3794 * __do_kmalloc - allocate memory
3795 * @size: how many bytes of memory are required.
3796 * @flags: the type of memory to allocate (see kmalloc).
3797 * @caller: function caller for debug tracking of the caller
3799 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3800 unsigned long caller
)
3802 struct kmem_cache
*cachep
;
3805 /* If you want to save a few bytes .text space: replace
3807 * Then kmalloc uses the uninlined functions instead of the inline
3810 cachep
= __find_general_cachep(size
, flags
);
3811 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3813 ret
= slab_alloc(cachep
, flags
, caller
);
3815 trace_kmalloc(caller
, ret
,
3816 size
, cachep
->size
, flags
);
3822 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3823 void *__kmalloc(size_t size
, gfp_t flags
)
3825 return __do_kmalloc(size
, flags
, _RET_IP_
);
3827 EXPORT_SYMBOL(__kmalloc
);
3829 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3831 return __do_kmalloc(size
, flags
, caller
);
3833 EXPORT_SYMBOL(__kmalloc_track_caller
);
3836 void *__kmalloc(size_t size
, gfp_t flags
)
3838 return __do_kmalloc(size
, flags
, 0);
3840 EXPORT_SYMBOL(__kmalloc
);
3844 * kmem_cache_free - Deallocate an object
3845 * @cachep: The cache the allocation was from.
3846 * @objp: The previously allocated object.
3848 * Free an object which was previously allocated from this
3851 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3853 unsigned long flags
;
3855 local_irq_save(flags
);
3856 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3857 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3858 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3859 __cache_free(cachep
, objp
, _RET_IP_
);
3860 local_irq_restore(flags
);
3862 trace_kmem_cache_free(_RET_IP_
, objp
);
3864 EXPORT_SYMBOL(kmem_cache_free
);
3867 * kfree - free previously allocated memory
3868 * @objp: pointer returned by kmalloc.
3870 * If @objp is NULL, no operation is performed.
3872 * Don't free memory not originally allocated by kmalloc()
3873 * or you will run into trouble.
3875 void kfree(const void *objp
)
3877 struct kmem_cache
*c
;
3878 unsigned long flags
;
3880 trace_kfree(_RET_IP_
, objp
);
3882 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3884 local_irq_save(flags
);
3885 kfree_debugcheck(objp
);
3886 c
= virt_to_cache(objp
);
3887 debug_check_no_locks_freed(objp
, c
->object_size
);
3889 debug_check_no_obj_freed(objp
, c
->object_size
);
3890 __cache_free(c
, (void *)objp
, _RET_IP_
);
3891 local_irq_restore(flags
);
3893 EXPORT_SYMBOL(kfree
);
3896 * This initializes kmem_list3 or resizes various caches for all nodes.
3898 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3901 struct kmem_list3
*l3
;
3902 struct array_cache
*new_shared
;
3903 struct array_cache
**new_alien
= NULL
;
3905 for_each_online_node(node
) {
3907 if (use_alien_caches
) {
3908 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3914 if (cachep
->shared
) {
3915 new_shared
= alloc_arraycache(node
,
3916 cachep
->shared
*cachep
->batchcount
,
3919 free_alien_cache(new_alien
);
3924 l3
= cachep
->nodelists
[node
];
3926 struct array_cache
*shared
= l3
->shared
;
3928 spin_lock_irq(&l3
->list_lock
);
3931 free_block(cachep
, shared
->entry
,
3932 shared
->avail
, node
);
3934 l3
->shared
= new_shared
;
3936 l3
->alien
= new_alien
;
3939 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3940 cachep
->batchcount
+ cachep
->num
;
3941 spin_unlock_irq(&l3
->list_lock
);
3943 free_alien_cache(new_alien
);
3946 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3948 free_alien_cache(new_alien
);
3953 kmem_list3_init(l3
);
3954 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3955 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3956 l3
->shared
= new_shared
;
3957 l3
->alien
= new_alien
;
3958 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3959 cachep
->batchcount
+ cachep
->num
;
3960 cachep
->nodelists
[node
] = l3
;
3965 if (!cachep
->list
.next
) {
3966 /* Cache is not active yet. Roll back what we did */
3969 if (cachep
->nodelists
[node
]) {
3970 l3
= cachep
->nodelists
[node
];
3973 free_alien_cache(l3
->alien
);
3975 cachep
->nodelists
[node
] = NULL
;
3983 struct ccupdate_struct
{
3984 struct kmem_cache
*cachep
;
3985 struct array_cache
*new[0];
3988 static void do_ccupdate_local(void *info
)
3990 struct ccupdate_struct
*new = info
;
3991 struct array_cache
*old
;
3994 old
= cpu_cache_get(new->cachep
);
3996 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3997 new->new[smp_processor_id()] = old
;
4000 /* Always called with the slab_mutex held */
4001 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
4002 int batchcount
, int shared
, gfp_t gfp
)
4004 struct ccupdate_struct
*new;
4007 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
4012 for_each_online_cpu(i
) {
4013 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
4016 for (i
--; i
>= 0; i
--)
4022 new->cachep
= cachep
;
4024 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
4027 cachep
->batchcount
= batchcount
;
4028 cachep
->limit
= limit
;
4029 cachep
->shared
= shared
;
4031 for_each_online_cpu(i
) {
4032 struct array_cache
*ccold
= new->new[i
];
4035 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4036 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4037 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4041 return alloc_kmemlist(cachep
, gfp
);
4044 /* Called with slab_mutex held always */
4045 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4051 * The head array serves three purposes:
4052 * - create a LIFO ordering, i.e. return objects that are cache-warm
4053 * - reduce the number of spinlock operations.
4054 * - reduce the number of linked list operations on the slab and
4055 * bufctl chains: array operations are cheaper.
4056 * The numbers are guessed, we should auto-tune as described by
4059 if (cachep
->size
> 131072)
4061 else if (cachep
->size
> PAGE_SIZE
)
4063 else if (cachep
->size
> 1024)
4065 else if (cachep
->size
> 256)
4071 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4072 * allocation behaviour: Most allocs on one cpu, most free operations
4073 * on another cpu. For these cases, an efficient object passing between
4074 * cpus is necessary. This is provided by a shared array. The array
4075 * replaces Bonwick's magazine layer.
4076 * On uniprocessor, it's functionally equivalent (but less efficient)
4077 * to a larger limit. Thus disabled by default.
4080 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4085 * With debugging enabled, large batchcount lead to excessively long
4086 * periods with disabled local interrupts. Limit the batchcount
4091 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4093 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4094 cachep
->name
, -err
);
4099 * Drain an array if it contains any elements taking the l3 lock only if
4100 * necessary. Note that the l3 listlock also protects the array_cache
4101 * if drain_array() is used on the shared array.
4103 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4104 struct array_cache
*ac
, int force
, int node
)
4108 if (!ac
|| !ac
->avail
)
4110 if (ac
->touched
&& !force
) {
4113 spin_lock_irq(&l3
->list_lock
);
4115 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4116 if (tofree
> ac
->avail
)
4117 tofree
= (ac
->avail
+ 1) / 2;
4118 free_block(cachep
, ac
->entry
, tofree
, node
);
4119 ac
->avail
-= tofree
;
4120 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4121 sizeof(void *) * ac
->avail
);
4123 spin_unlock_irq(&l3
->list_lock
);
4128 * cache_reap - Reclaim memory from caches.
4129 * @w: work descriptor
4131 * Called from workqueue/eventd every few seconds.
4133 * - clear the per-cpu caches for this CPU.
4134 * - return freeable pages to the main free memory pool.
4136 * If we cannot acquire the cache chain mutex then just give up - we'll try
4137 * again on the next iteration.
4139 static void cache_reap(struct work_struct
*w
)
4141 struct kmem_cache
*searchp
;
4142 struct kmem_list3
*l3
;
4143 int node
= numa_mem_id();
4144 struct delayed_work
*work
= to_delayed_work(w
);
4146 if (!mutex_trylock(&slab_mutex
))
4147 /* Give up. Setup the next iteration. */
4150 list_for_each_entry(searchp
, &slab_caches
, list
) {
4154 * We only take the l3 lock if absolutely necessary and we
4155 * have established with reasonable certainty that
4156 * we can do some work if the lock was obtained.
4158 l3
= searchp
->nodelists
[node
];
4160 reap_alien(searchp
, l3
);
4162 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4165 * These are racy checks but it does not matter
4166 * if we skip one check or scan twice.
4168 if (time_after(l3
->next_reap
, jiffies
))
4171 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4173 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4175 if (l3
->free_touched
)
4176 l3
->free_touched
= 0;
4180 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4181 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4182 STATS_ADD_REAPED(searchp
, freed
);
4188 mutex_unlock(&slab_mutex
);
4191 /* Set up the next iteration */
4192 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4195 #ifdef CONFIG_SLABINFO
4196 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4199 unsigned long active_objs
;
4200 unsigned long num_objs
;
4201 unsigned long active_slabs
= 0;
4202 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4206 struct kmem_list3
*l3
;
4210 for_each_online_node(node
) {
4211 l3
= cachep
->nodelists
[node
];
4216 spin_lock_irq(&l3
->list_lock
);
4218 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4219 if (slabp
->inuse
!= cachep
->num
&& !error
)
4220 error
= "slabs_full accounting error";
4221 active_objs
+= cachep
->num
;
4224 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4225 if (slabp
->inuse
== cachep
->num
&& !error
)
4226 error
= "slabs_partial inuse accounting error";
4227 if (!slabp
->inuse
&& !error
)
4228 error
= "slabs_partial/inuse accounting error";
4229 active_objs
+= slabp
->inuse
;
4232 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4233 if (slabp
->inuse
&& !error
)
4234 error
= "slabs_free/inuse accounting error";
4237 free_objects
+= l3
->free_objects
;
4239 shared_avail
+= l3
->shared
->avail
;
4241 spin_unlock_irq(&l3
->list_lock
);
4243 num_slabs
+= active_slabs
;
4244 num_objs
= num_slabs
* cachep
->num
;
4245 if (num_objs
- active_objs
!= free_objects
&& !error
)
4246 error
= "free_objects accounting error";
4248 name
= cachep
->name
;
4250 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4252 sinfo
->active_objs
= active_objs
;
4253 sinfo
->num_objs
= num_objs
;
4254 sinfo
->active_slabs
= active_slabs
;
4255 sinfo
->num_slabs
= num_slabs
;
4256 sinfo
->shared_avail
= shared_avail
;
4257 sinfo
->limit
= cachep
->limit
;
4258 sinfo
->batchcount
= cachep
->batchcount
;
4259 sinfo
->shared
= cachep
->shared
;
4260 sinfo
->objects_per_slab
= cachep
->num
;
4261 sinfo
->cache_order
= cachep
->gfporder
;
4264 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4268 unsigned long high
= cachep
->high_mark
;
4269 unsigned long allocs
= cachep
->num_allocations
;
4270 unsigned long grown
= cachep
->grown
;
4271 unsigned long reaped
= cachep
->reaped
;
4272 unsigned long errors
= cachep
->errors
;
4273 unsigned long max_freeable
= cachep
->max_freeable
;
4274 unsigned long node_allocs
= cachep
->node_allocs
;
4275 unsigned long node_frees
= cachep
->node_frees
;
4276 unsigned long overflows
= cachep
->node_overflow
;
4278 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4279 "%4lu %4lu %4lu %4lu %4lu",
4280 allocs
, high
, grown
,
4281 reaped
, errors
, max_freeable
, node_allocs
,
4282 node_frees
, overflows
);
4286 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4287 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4288 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4289 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4291 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4292 allochit
, allocmiss
, freehit
, freemiss
);
4297 #define MAX_SLABINFO_WRITE 128
4299 * slabinfo_write - Tuning for the slab allocator
4301 * @buffer: user buffer
4302 * @count: data length
4305 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4306 size_t count
, loff_t
*ppos
)
4308 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4309 int limit
, batchcount
, shared
, res
;
4310 struct kmem_cache
*cachep
;
4312 if (count
> MAX_SLABINFO_WRITE
)
4314 if (copy_from_user(&kbuf
, buffer
, count
))
4316 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4318 tmp
= strchr(kbuf
, ' ');
4323 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4326 /* Find the cache in the chain of caches. */
4327 mutex_lock(&slab_mutex
);
4329 list_for_each_entry(cachep
, &slab_caches
, list
) {
4330 if (!strcmp(cachep
->name
, kbuf
)) {
4331 if (limit
< 1 || batchcount
< 1 ||
4332 batchcount
> limit
|| shared
< 0) {
4335 res
= do_tune_cpucache(cachep
, limit
,
4342 mutex_unlock(&slab_mutex
);
4348 #ifdef CONFIG_DEBUG_SLAB_LEAK
4350 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4352 mutex_lock(&slab_mutex
);
4353 return seq_list_start(&slab_caches
, *pos
);
4356 static inline int add_caller(unsigned long *n
, unsigned long v
)
4366 unsigned long *q
= p
+ 2 * i
;
4380 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4386 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4392 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4393 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4395 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4400 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4402 #ifdef CONFIG_KALLSYMS
4403 unsigned long offset
, size
;
4404 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4406 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4407 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4409 seq_printf(m
, " [%s]", modname
);
4413 seq_printf(m
, "%p", (void *)address
);
4416 static int leaks_show(struct seq_file
*m
, void *p
)
4418 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4420 struct kmem_list3
*l3
;
4422 unsigned long *n
= m
->private;
4426 if (!(cachep
->flags
& SLAB_STORE_USER
))
4428 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4431 /* OK, we can do it */
4435 for_each_online_node(node
) {
4436 l3
= cachep
->nodelists
[node
];
4441 spin_lock_irq(&l3
->list_lock
);
4443 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4444 handle_slab(n
, cachep
, slabp
);
4445 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4446 handle_slab(n
, cachep
, slabp
);
4447 spin_unlock_irq(&l3
->list_lock
);
4449 name
= cachep
->name
;
4451 /* Increase the buffer size */
4452 mutex_unlock(&slab_mutex
);
4453 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4455 /* Too bad, we are really out */
4457 mutex_lock(&slab_mutex
);
4460 *(unsigned long *)m
->private = n
[0] * 2;
4462 mutex_lock(&slab_mutex
);
4463 /* Now make sure this entry will be retried */
4467 for (i
= 0; i
< n
[1]; i
++) {
4468 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4469 show_symbol(m
, n
[2*i
+2]);
4476 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4478 return seq_list_next(p
, &slab_caches
, pos
);
4481 static void s_stop(struct seq_file
*m
, void *p
)
4483 mutex_unlock(&slab_mutex
);
4486 static const struct seq_operations slabstats_op
= {
4487 .start
= leaks_start
,
4493 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4495 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4498 ret
= seq_open(file
, &slabstats_op
);
4500 struct seq_file
*m
= file
->private_data
;
4501 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4510 static const struct file_operations proc_slabstats_operations
= {
4511 .open
= slabstats_open
,
4513 .llseek
= seq_lseek
,
4514 .release
= seq_release_private
,
4518 static int __init
slab_proc_init(void)
4520 #ifdef CONFIG_DEBUG_SLAB_LEAK
4521 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4525 module_init(slab_proc_init
);
4529 * ksize - get the actual amount of memory allocated for a given object
4530 * @objp: Pointer to the object
4532 * kmalloc may internally round up allocations and return more memory
4533 * than requested. ksize() can be used to determine the actual amount of
4534 * memory allocated. The caller may use this additional memory, even though
4535 * a smaller amount of memory was initially specified with the kmalloc call.
4536 * The caller must guarantee that objp points to a valid object previously
4537 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4538 * must not be freed during the duration of the call.
4540 size_t ksize(const void *objp
)
4543 if (unlikely(objp
== ZERO_SIZE_PTR
))
4546 return virt_to_cache(objp
)->object_size
;
4548 EXPORT_SYMBOL(ksize
);